Intelligent fresh-keeping system and method for food refrigeration equipment

Abstract

The present invention discloses an intelligent fresh-keeping system and method for food refrigeration equipment, comprising: firstly, collecting an initial image of a target food through an image sensor; then processing the image through a food category recognition model to obtain the appearance and type of the food to be monitored, and determining a key monitoring area; finally, collecting image data of the key area, combining the food type matching preset library to obtain the current fresh-keeping strategy and sending it to a central control chip for execution, so as to realize intelligent and accurate fresh-keeping of food.

Description

Technical Field

[0001] The present invention relates to the field of food preservation, and in particular to an intelligent fresh-keeping system and method for food refrigeration equipment. Background Art

[0002] In the field of food storage, ensuring the freshness of food is of vital importance. Traditional food refrigeration equipment is difficult to accurately preserve food according to the specific conditions. With the development of science and technology, there is an urgent need for intelligent preservation. Existing technologies are insufficient in accurately identifying food characteristics and formulating personalized fresh-keeping strategies based on them. They cannot effectively meet the fresh-keeping requirements of diversified foods during refrigeration storage, which can easily lead to problems such as food spoilage. Therefore, a system that can intelligently and accurately lock freshness is needed. Summary of the invention

[0003] The object of the present invention is to provide an intelligent fresh-keeping system and method for food refrigeration equipment.

[0004] In a first aspect, an embodiment of the present invention provides an intelligent fresh-keeping system for food refrigeration equipment, comprising:

[0005] An acquisition module is used to acquire an image of a target food through an image sensor of a food refrigeration device to obtain an initial food image;

[0006] A recognition module, used to input the initial food image into a pre-trained food category recognition model for processing, to obtain the appearance and food type of the target food to be monitored; based on the appearance of the food to be monitored, to determine at least one key monitoring area;

[0007] The fresh-keeping module is used to collect image data of at least one key monitoring area, and match the food type in a preset fresh-keeping strategy library to obtain a current fresh-keeping strategy for the target food; and send the current fresh-keeping strategy to the central control chip of the food refrigeration equipment to execute the current fresh-keeping strategy.

[0008] In a second aspect, an embodiment of the present invention provides an intelligent fresh-keeping method for food refrigeration equipment, comprising:

[0009] Capturing an image of the target food through an image sensor of the food refrigeration equipment to obtain an initial food image;

[0010] Inputting the initial food image into a pre-trained food category recognition model for processing to obtain the appearance and food type of the target food to be monitored;

[0011] Based on the appearance of the food to be monitored, determining at least one key monitoring area;

[0012] Collecting image data of the at least one key monitoring area, and matching the image data in a preset fresh-keeping strategy library in combination with the food type, to obtain a current fresh-keeping strategy for the target food;

[0013] The current fresh-keeping strategy is sent to the central control chip of the food refrigeration equipment to execute the current fresh-keeping strategy.

[0014] Compared with the prior art, the beneficial effects provided by the present invention include: adopting the intelligent fresh-keeping system and method for food refrigeration equipment disclosed in the present invention, collecting the initial image of the target food through the image sensor; then processing the appearance and type of the food to be monitored through the food category recognition model, and determining the key monitoring area; finally collecting the image data of the key area, combining the food type matching preset library to obtain the current fresh-keeping strategy and sending it to the central control chip for execution, so as to realize intelligent and precise fresh-keeping of food. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative work.

[0016] Figure 1 A schematic flow chart of the steps of an intelligent fresh-keeping method for food refrigeration equipment provided by an embodiment of the present invention;

[0017] Figure 2 A schematic block diagram of the structure of an intelligent fresh-keeping system for food refrigeration equipment provided by an embodiment of the present invention;

[0018] Figure 3 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0020] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0021] In order to solve the technical problems in the aforementioned background technology, Figure 1This is a flow chart of an intelligent fresh-keeping method for food refrigeration equipment provided in an embodiment of the present disclosure. The intelligent fresh-keeping method for food refrigeration equipment is introduced in detail below.

[0022] Step S201, capturing an image of a target food through an image sensor of a food refrigeration device to obtain an initial food image;

[0023] Step S202, inputting the initial food image into a pre-trained food category recognition model for processing to obtain the food appearance and food type of the target food to be monitored;

[0024] Step S203, determining at least one key monitoring area based on the appearance of the food to be monitored;

[0025] Step S204, collecting image data of the at least one key monitoring area, and matching the image data in a preset fresh-keeping strategy library in combination with the food type to obtain a current fresh-keeping strategy for the target food;

[0026] Step S205: sending the current fresh-keeping strategy to the central control chip of the food refrigeration equipment to execute the current fresh-keeping strategy.

[0027] In an embodiment of the present invention, for example, assume that in the fresh food area of ​​a large supermarket, there are many different kinds of food placed in various food refrigeration equipment (such as refrigerators, freezers, etc.) for preservation and storage. In order to better ensure the freshness of food, the supermarket uses food refrigeration equipment with an intelligent fresh-keeping system, and the server, as the core execution body of the entire intelligent fresh-keeping process, is responsible for processing relevant data and decisions. A batch of fresh apples are stored in the refrigerator, and the server controls the image sensor in the refrigerator to start the collection work. The image sensor is installed in a suitable position inside the refrigerator, and can clearly capture the apples placed in its corresponding area. The sensor takes pictures of the apples at a certain time interval (for example, every 30 minutes), and the photos taken each time are the initial food images of these apples. These images will be transmitted to the server in real time for subsequent processing. For example, the initial food image obtained by this shooting clearly shows the overall shape, color, and some tiny features that may exist on the surface of the apple, etc., which provides basic visual information for the subsequent identification of the state of the apple. After the server receives the initial food image of the apple transmitted from the image sensor of the refrigerator, it immediately inputs it into the pre-trained food category recognition model. This food category recognition model is trained with a large amount of different types of food image data, and can accurately identify the categories of many common foods and their appearance features. For the input apple initial food image, the model will first perform a series of feature extraction operations on the image. It will identify the appearance features of the apple, such as the round outline, red or green skin color, and possible spots on the skin. At the same time, through the classification mechanism within the model, it is clearly determined that these target foods are of the food type "apple". After being processed by the model, the server obtains the appearance information of the food to be monitored for this batch of apples (such as the various appearance features mentioned above) and the clear food type (apple). After obtaining the appearance information of the food to be monitored for the apples, the server begins to determine the key monitoring area. Considering that the most likely place for apples to deteriorate during storage may be the part where the skin is damaged by bumps or the connection of the fruit stem. Based on the appearance features of the apples obtained previously, such as some tiny bump marks on the skin, the area where these bump marks are located will be focused on and may be determined as a key monitoring area. In addition, the fruit stem connection is also marked as a key monitoring area because it is easy to breed bacteria and cause the apple to spoil. Through a comprehensive analysis of the appearance of the apple food to be monitored, the server finally determined several key monitoring areas, including local epidermal areas with obvious bumps and a small area around the fruit stem connection, so that changes in these parts can be monitored more accurately in the future. The server controls the image sensor in the refrigerator to start again, this time specifically for the key monitoring areas of the apples previously determined to collect image data.The sensor will capture these key areas at a higher resolution and more frequent intervals (for example, every 15 minutes) to obtain more detailed image data, which will be fed back to the server in real time. At the same time, the server will search and match the preset fresh-keeping strategy library in combination with the food type "apple" that has been determined. This fresh-keeping strategy library stores various fresh-keeping strategies for different food types. For apples, there may be different fresh-keeping strategies formulated according to different storage stages (such as fresh apples that have just been picked, apples that have been stored for a certain period of time but are still fresh, and apples that are beginning to show signs of slight deterioration, etc.). Assuming that the server determines that this batch of apples has been stored for a period of time but is still fresh based on the collected key area image data, the current fresh-keeping strategy matched in the fresh-keeping strategy library may be: appropriately lower the temperature in the refrigerator to a certain value (such as 3°C), and slightly increase the air humidity to a certain proportion (such as 85%) to extend the shelf life of the apples. After determining the current fresh-keeping strategy for this batch of apples, the server will quickly send the strategy information to the central control chip of the refrigerator in a specific data format. After receiving the instructions from the server, the central control chip will perform corresponding control operations on the refrigerator's refrigeration system, humidity control system, etc. according to the strategy requirements. For example, the central control chip will control the refrigeration system to gradually reduce the temperature in the refrigerator to 3°C, and at the same time command the humidity control system to increase the air humidity to 85%, thereby realizing the intelligent fresh-keeping operation of apples and ensuring that the apples can remain fresh in the refrigerator for as long as possible. Through the above detailed scenario examples of each step with the server as the execution body, it can be clearly seen how the intelligent fresh-keeping method of food refrigeration equipment operates in actual scenarios to achieve effective monitoring and control of food preservation.

[0028] In the embodiment of the present invention, the determining of at least one key monitoring area based on the appearance of the food to be monitored may be implemented through the following examples.

[0029] Acquire the appearance of a food to be monitored and food deterioration description information corresponding to the appearance of the food to be monitored, wherein the appearance of the food to be monitored includes at least one food appearance image;

[0030] Performing feature extraction processing on the food appearance image to obtain at least one appearance feature of the evaluation attribute, and performing feature extraction processing on the food deterioration description information to obtain monitoring content features;

[0031] Determine a feature matching coefficient between the appearance feature and the monitoring content feature in each image grid unit, wherein the feature matching coefficient is used to reflect the matching situation between the appearance feature and the monitoring content feature in the same image grid unit;

[0032] Based on the feature matching coefficient, performing a feature integration operation on the appearance feature and the monitoring content feature to obtain an appearance detection feature;

[0033] At least one key monitoring area is determined in the food appearance image based on the appearance detection feature and the appearance feature.

[0034] In the embodiment of the present invention, for example, still taking the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server has previously obtained the relevant information such as the appearance of the food to be monitored of the apples, and now it is necessary to further determine the key monitoring area. The server obtains the appearance information of the food to be monitored about the apples from the previous processing flow, which includes multiple images of the appearance of apples taken from different angles, which can clearly present the overall shape and color of the apples. At the same time, the server also stores a description of food spoilage related to apples, which records in detail the various characteristic changes that may occur in the apples during the spoilage process, such as the darkening of the skin color, the appearance of soft and rotten spots, and the blackening of the fruit stems. For the acquired apple appearance image, the server uses a specific image processing algorithm to perform feature extraction processing on it. For example, for the circular contour of the apple, the algorithm will extract its shape features; for the color of the apple, it will extract features such as chroma and saturation, which are classified as appearance features of different evaluation attributes. At the same time, the server also performs feature extraction processing on the description information of food spoilage of the apple. For example, for the description of "skin color darkens", the monitoring content features such as the degree and range of color change are extracted; for "soft and rotten spots appear", the features such as the size and distribution density of the spots are extracted, thereby obtaining a series of monitoring content features. The server divides the apple appearance image into multiple small image grid units. Then, for each image grid unit, the appearance features therein are compared with the corresponding monitoring content features. For example, in a certain grid unit, the apple appearance image shows that the skin color here has slightly changed, and the monitoring content feature of the skin color darkening mentioned in the food spoilage description information is related to it. Through complex calculations, the server comprehensively considers factors such as the degree and range of color change, and determines the feature matching coefficient under the grid unit to accurately reflect the matching situation of the appearance feature and the monitoring content feature here. The server performs weighted integration operations on the corresponding appearance features and monitoring content features based on the feature matching coefficients under each image grid unit. For example, for a grid unit with a high feature matching coefficient, the corresponding appearance features and monitoring content features have relatively high weights when integrated. After such integration processing, an appearance detection feature that comprehensively considers the appearance features and the deterioration monitoring situation is obtained. The server analyzes the appearance detection features obtained and the initial appearance features. For areas that show a high match with deterioration characteristics in the appearance detection features, for example, the color change characteristics of a local area are highly matched with the deterioration description feature of "darkening of the skin color", and there are some previously observed conditions in the appearance characteristics of this area, such as slight bumps, then the server will identify this area as a key monitoring area so that it can focus on monitoring the changes in the apples here in the future.

[0035] In the embodiment of the present invention, the determination of the feature matching coefficient between the appearance feature and the monitoring content feature in each image grid unit may be implemented through the following example.

[0036] Aligning the appearance features and the monitoring content features into the same vector dimension space to obtain target appearance features of each evaluation attribute and target monitoring content features corresponding to the target appearance features;

[0037] The local image features corresponding to each image grid unit are collected from the target appearance features, and the feature matching coefficients between the local image features and the target monitoring content features are determined.

[0038] In the embodiment of the present invention, illustratively, continuing to take the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server has obtained the appearance features of the apples and the monitoring content features extracted from the corresponding food spoilage description information, and now it is necessary to determine their feature matching coefficients under each image grid unit. After the server receives the appearance features and monitoring content features of the apples, in order to more accurately compare and analyze them, it is necessary to align them to the same vector dimension space. For example, the appearance features of the apples contain feature information of multiple attributes such as shape (round), color (red and some areas are slightly lighter in color), and the monitoring content features have features such as skin color changes (such as the range of reddening and darkening), whether there are soft and rotten spots (spot size, distribution density, etc.). The server processes these features of different types and dimensions through a specific algorithm, and adjusts their attribute dimensions separately, so that each evaluation attribute of the appearance feature (such as shape, color, etc.) and each monitoring indicator feature of the monitoring content feature (such as color change degree, spot situation, etc.) can be reasonably represented in a unified vector dimension space. After this process, the target appearance features of each evaluation attribute and the corresponding target monitoring content features are obtained. For example, after alignment, the color features of apples can reflect the specific conditions of their color in this space in a more standardized and unified numerical form, and can be compared and analyzed with the features of the skin color changes in the monitoring content features under the same dimensional standard. The server divides the apple appearance image corresponding to the aligned target appearance features into many small image grid units. For example, divide an overall appearance image of an apple into small square grid units with a side length of several millimeters. Then, for each image grid unit, the server collects the local image features corresponding to the unit from the target appearance features. For example, in a grid unit near the top of the apple, the collected local image features may be the local color of the apple skin here (slightly lighter than the surrounding area) and the local shape outline (slightly irregular, possibly due to growth). Then, the server compares and analyzes this collected local image feature with the target monitoring content feature. Taking the grid unit located near the top of the apple as an example, the server will compare the local color conditions collected here with the characteristics of the skin color changes in the monitoring content features, consider whether the degree and range of the color changes are consistent, as well as factors such as the local shape contour and the shape change characteristics that may be caused by deterioration. Through a series of complex calculations and evaluations, the server will ultimately determine the feature matching coefficient between the local image features of this image grid unit and the target monitoring content features, thereby accurately reflecting the matching between the appearance features and the monitoring content features at this grid unit.

[0039] In an embodiment of the present invention, the monitoring content feature includes at least one monitoring indicator feature, and the appearance feature and the monitoring content feature are aligned in the same vector dimension space to obtain the target appearance feature of each evaluation attribute and the target monitoring content feature corresponding to the target appearance feature. The package can be implemented through the following examples.

[0040] Identifying feature attribute dimensions in the appearance features to obtain appearance attribute dimensions, and identifying feature attribute dimensions in the monitored content features to obtain content attribute dimensions;

[0041] Based on the appearance attribute dimension and the content attribute dimension, determining the target attribute dimension of the vector dimension space corresponding to each evaluation attribute, and performing a feature integration operation on the monitoring indicator feature to obtain an integrated monitoring content feature;

[0042] The attribute dimensions of the appearance feature and the attribute dimensions of the integrated monitoring content feature are respectively aligned to the target attribute dimension to obtain the target appearance feature of each evaluation attribute and the target monitoring content feature corresponding to the target appearance feature.

[0043] In the embodiment of the present invention, for example, taking the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server has obtained the appearance features of the apples and the corresponding monitoring content features containing multiple monitoring indicator features, and now it is necessary to align them to the same vector dimension space. The server obtains the appearance feature information of the apples, such as the appearance features of the apples, including shape (circular), color (red and with color differences in different areas), surface smoothness, etc. For the characteristic attribute of shape, the server analyzes it through a specific algorithm and identifies that it may be described by data of several dimensions, such as using information such as the coordinates of the center of the circle and the radius to determine the circular shape. The number of dimensions identified here for describing the shape feature attribute is the appearance attribute dimension, assuming that the shape feature attribute dimension is 3. At the same time, for the monitoring content features, for example, the monitoring indicator features of apple deterioration include changes in skin color (described by the range of color value changes, etc.), whether there are soft and rotten spots (described by the number, size, distribution, etc. of spots), etc. The server analyzes the monitoring indicator feature of the skin color change and identifies the number of dimensions required to describe it, such as using several dimensions such as hue, saturation, and brightness range to describe it. This is the content attribute dimension. Assume that the content attribute dimension of the skin color change is 3. The server comprehensively considers the attribute dimensions of the previously identified apple appearance features (such as 3 dimensions of shape, and other dimensions of color, etc.) and the attribute dimensions of each monitoring indicator feature in the monitoring content feature (such as 3 dimensions of skin color change, etc.). For the evaluation attribute of the shape of the apple, through analysis and comparison, the appropriate target attribute dimension in the vector dimension space is determined. Assume that after comprehensive consideration, the target attribute dimension corresponding to the shape evaluation attribute is still 3. Then, the server performs feature integration operations on each monitoring indicator feature in the monitoring content feature. For example, for the two monitoring indicator features of the skin color change and whether there are soft and rotten spots in the apple deterioration monitoring content feature, the server integrates them according to certain rules through a specific algorithm, and combines the relevant information about the color change and the spot situation in a more comprehensive and easier way to process later, and obtains the integrated monitoring content feature. The server aligns and adjusts the attribute dimension of the shape feature in the apple appearance feature (previously assumed to be 3) according to the determined target attribute dimension (the shape evaluation attribute also corresponds to 3). It is possible to normalize the specific data describing the shape (center coordinates, radius, etc.) so that it can be accurately represented in the vector dimension space that conforms to the target attribute dimension, thereby obtaining the target appearance feature of the shape evaluation attribute. Similarly, the server aligns the attribute dimension of the integrated monitoring content features to the target attribute dimension. For example, for the integrated monitoring content features that include epidermal color changes and soft and rotten spots, by adjusting the representation of the relevant data, etc., it can obtain the target monitoring content features corresponding to the shape evaluation attribute in the same vector dimension space as the target appearance features.In this way, the target appearance features and target monitoring content features for each evaluation attribute are acquired so as to facilitate more accurate analysis and processing in the future.

[0044] In the embodiment of the present invention, the determination of the feature matching coefficient between the local image feature and the target monitoring content feature may be implemented through the following example.

[0045] Collecting features under each feature attribute in the local image features to obtain appearance attribute features;

[0046] Extracting features under the feature attributes corresponding to the appearance attribute features from the target monitoring content features to obtain content attribute features;

[0047] Based on the appearance attribute feature and the content attribute feature, a feature matching coefficient between the local image feature and the target monitoring content feature is determined.

[0048] In the embodiment of the present invention, illustratively, still taking the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server has obtained the local image features of each image grid unit corresponding to the target appearance features of the apple in the previous operation, as well as the corresponding target monitoring content features, and now it is necessary to determine the feature matching coefficient between them. The server analyzes and processes the local image features of a certain image grid unit. For example, the local image feature corresponds to a small area of ​​the apple skin, and this local image feature contains information such as shape, color, texture, etc. For the shape feature attribute, the server accurately collects the specific shape characteristics of the apple skin in this area through a specific image analysis algorithm, such as being approximately circular but with slightly irregular edges, which is the feature under the shape feature attribute; for the color feature attribute, the actual color of the skin here is collected, such as showing light red and slightly yellowing in some parts; for the texture feature attribute, the fineness and direction of the skin texture can be identified. By summarizing these features collected from different feature attributes, the appearance attribute features corresponding to the local image feature are obtained. The server then searches and extracts in the existing target monitoring content features. The target monitoring content feature is a comprehensive feature description of the possible deterioration of apples after integration. Assuming that the previously determined appearance attribute features include color features (such as light red and slightly yellow in some areas), the server will extract color-related features corresponding to the deterioration of apples in the target monitoring content features. For example, it may be extracted that the color of the apple skin will gradually deepen and become partially dark in the early stage of deterioration; for shape, if the appearance attribute features mention that the edges are slightly irregular, the target monitoring content features may extract shape-related features such as the apple may change shape and become shriveled due to water loss during the deterioration process; for texture, it may extract content attribute features such as the roughness of the skin texture during deterioration. The features extracted from the target monitoring content features corresponding to the appearance attribute features are summarized to obtain the content attribute features. The server compares and analyzes the acquired appearance attribute features and content attribute features. Taking the color feature as an example, the light red and slightly yellow color collected from the local image feature is compared with the color of the apple skin that gradually deepens and becomes partially dark in the early stage of deterioration extracted from the target monitoring content feature. Through a series of complex calculations, taking into account factors such as the direction of color change (darker or lighter), degree (magnitude of change), and range (local or large area), and combining the comparison results of various feature attributes (shape, color, texture, etc.), the feature matching coefficient between the local image feature and the target monitoring content feature is finally determined, so as to accurately reflect the matching situation between the local image feature and the target monitoring content feature at the image grid unit, so as to facilitate further analysis and processing based on this.

[0049] In the embodiment of the present invention, the feature matching coefficient between the local image feature and the target monitoring content feature is determined based on the appearance attribute feature and the monitoring content feature attribute, which can be implemented through the following examples.

[0050] Performing a feature integration operation on the appearance attribute feature and the content attribute feature to obtain an integrated attribute feature corresponding to each feature attribute;

[0051] Superimposing the integrated attribute features under each image grid unit to obtain a first integrated feature corresponding to each image grid unit;

[0052] A feature integration operation is performed on the local image feature and the target monitoring content feature to obtain a second integrated feature, and a relative ratio between the second integrated feature and the first integrated feature is determined to obtain a feature matching coefficient corresponding to each image grid unit.

[0053] In the embodiment of the present invention, illustratively, taking the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server has obtained the appearance attribute features corresponding to the local image features of the specific image grid unit, and the content attribute features extracted from the target monitoring content features, and now the feature matching coefficient is determined based on these. The server processes the appearance attribute features (such as the shape, color, texture and other aspects of the apple skin in the area) and content attribute features (such as the change features that may occur in these aspects when the corresponding apple deteriorates) of a certain image grid unit of the apple. For the characteristic attribute of shape, if the appearance attribute feature is that the skin of the area is approximately circular but the edges are slightly irregular, and the content attribute feature is that the shape may become shriveled due to water loss when the apple deteriorates, the server integrates these two shape features according to certain rules through a specific algorithm, comprehensively considers the current actual shape and the possible shape change trend after deterioration, and obtains the integrated attribute features corresponding to the characteristic attribute of shape. Similarly, other characteristic attributes such as color and texture are also subjected to such integration operations, so as to obtain the integrated attribute features corresponding to each characteristic attribute. After completing the above integration operation for each feature attribute, the server will perform superposition processing on all the integrated attribute features under the image grid unit. For example, for the apple image grid unit just processed, the integrated attribute features corresponding to the feature attributes such as shape, color, and texture have been obtained. The server superimposes the integrated attribute features of these different feature attributes according to a specific calculation method, such as simple numerical addition (if the feature is expressed in numerical form) or weighted addition according to a certain weight (different weights are assigned according to the importance of different feature attributes). Through such a superposition operation, the first integrated feature corresponding to the image grid unit is finally obtained. This first integrated feature integrates the information of each feature attribute under the grid unit after integration. The server then performs feature integration operation again on the original local image features of the image grid unit (including initial information of various aspects such as shape, color, texture, etc.) and the target monitoring content features (a comprehensive description of the deterioration of apples). Through a specific algorithm, these two parts of information are comprehensively combined to obtain the second integrated feature. Then, the server determines the feature matching coefficient by calculating the relative ratio between the second integrated feature and the first integrated feature obtained previously. For example, if the value of the second integrated feature is twice the value of the first integrated feature, then the feature matching coefficient may be 2 (the specific calculation method will be determined according to the actual algorithm used). This feature matching coefficient accurately reflects the matching between the local image features and the target monitoring content features under the image grid unit, providing an important basis for subsequent analysis and processing.

[0054] In the embodiment of the present invention, based on the feature matching coefficient, the appearance feature and the monitoring content feature are subjected to a feature integration operation to obtain an appearance detection feature, which can be implemented through the following example.

[0055] Based on the feature matching coefficient, the local image features are weighted to obtain original appearance detection features corresponding to each image grid unit;

[0056] A feature integration operation is performed on the original appearance detection features of the same evaluation attribute to obtain appearance detection features of each evaluation attribute.

[0057] In the embodiment of the present invention, illustratively, continuing to take the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server has determined the feature matching coefficients between the local image features and the target monitoring content features under each image grid unit in the previous step, and now it is necessary to perform subsequent feature integration operations based on these coefficients to obtain appearance detection features. The server obtains the local image features of a certain image grid unit of the apple, for example, the local image features contain the specific conditions of the color, texture, shape, etc. of the apple skin in this area. At the same time, the feature matching coefficient corresponding to the image grid unit has also been clarified. Assume that the feature matching coefficient of this image grid unit is 1.5 (indicating that the local image features and the target monitoring content features have a certain degree of matching here). The server will perform weighted processing on the local image features according to this coefficient. For the color feature, if the original color feature value of the apple skin in this area is (assuming) 80 (here is just an example of a possible quantitative representation), then the weighted color feature value becomes 80×1.5=120 (the specific calculation method depends on the actual weighting algorithm used). The various aspects of other local image features such as texture and shape are also weighted according to this coefficient. After such weighted processing, the original appearance detection features corresponding to this image grid unit are obtained, which comprehensively considers the local image features and the matching with the target monitoring content features. After the server completes the acquisition of the original appearance detection features of each image grid unit, it starts to integrate the original appearance detection features of the same evaluation attribute. For example, for the evaluation attribute of apple skin color, there may be multiple image grid units that involve color-related original appearance detection features. Assume that in three different image grid units, the color-related original appearance detection features obtained after the previous weighted processing are 120 (the weighted value in the previous example), 130 and 110 respectively. The server will integrate these three original appearance detection features belonging to the color evaluation attribute through a specific algorithm. It may be a simple numerical addition and then take the average value (it may also be integrated according to a more complex algorithm, depending on the actual situation). Assuming that the method of taking the average value after addition is adopted here, the appearance detection feature of the color evaluation attribute is (120+130+110)÷3=120 (here is just an example to illustrate a possible integration result). Similarly, for other evaluation attributes such as the shape and texture of the apple skin, the corresponding original appearance detection features are also integrated in this way, and finally the appearance detection features of each evaluation attribute are obtained. These appearance detection features can more comprehensively and accurately reflect the relationship between the appearance of the apple and the possible deterioration, and provide an important basis for subsequent operations such as determining key monitoring areas in food appearance images.

[0058] In the embodiment of the present invention, determining at least one key monitoring area in the food appearance image based on the appearance detection feature and the appearance feature can be implemented through the following examples.

[0059] Based on the evaluation attribute of the appearance feature, performing a saliency evaluation on the appearance feature, and extracting the appearance feature of the target evaluation attribute from the appearance feature based on the saliency evaluation result to obtain a candidate appearance feature;

[0060] Extracting the appearance detection features corresponding to the target evaluation attributes from the appearance detection features to obtain candidate appearance detection features;

[0061] According to the alternative appearance detection features and the alternative appearance features, monitoring area features are collected in the food appearance image, and based on the monitoring area features, at least one key monitoring area in the food appearance image is determined.

[0062] In the embodiment of the present invention, illustratively, still taking the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server has obtained the appearance detection features and appearance features of the apples, and now it is necessary to determine the key monitoring areas in the food appearance image based on these. The server obtains the appearance features of the apples, which include evaluation attributes such as shape (mainly round, with slightly irregular edges in some parts), color (overall red, with light differences in some parts), and surface texture (relatively delicate). For the evaluation attribute of color, the server will analyze its prominence in the entire apple appearance. For example, if it is found that there is an area on the surface of the apple that is obviously darker than other places, this has a higher significance in terms of color attributes. Through similar analysis of each evaluation attribute (shape, color, texture, etc.), the significance of each evaluation attribute is determined. Based on the above-mentioned significance evaluation results, the server extracts the appearance features of those target evaluation attributes with higher significance. Assuming that the dark color area is more prominent in the significance evaluation, the color features corresponding to the area and the possibly associated shape, texture and other features are extracted as alternative appearance features. The server then searches in the existing appearance detection features. The appearance detection feature is the result of the integration of the appearance feature and the monitoring content feature, and is associated with the possible deterioration of the apple. For the target evaluation attribute (such as the color mentioned above) that was concerned when extracting the alternative appearance features, find the corresponding part in the appearance detection feature. For example, in the appearance detection feature, the feature part related to the dark color area that can reflect the possible deterioration trend of the area is extracted to obtain the alternative appearance detection feature. The server combines the alternative appearance detection feature with the alternative appearance feature for analysis. For example, the alternative appearance feature shows that a certain area of ​​the apple is dark in color and slightly deformed in shape. The alternative appearance detection feature indicates that there is a possible deterioration trend in the area (such as the color change conforms to a certain deterioration pattern). Based on this information, the server accurately locates the area in the food appearance image and collects the features of the area and a certain range around it as the monitoring area features. This may include more detailed color changes, texture changes, etc. Through further analysis of the collected monitoring area characteristics, if it is found that the change trend in the area is continuous and conforms to the law of possible deterioration, the server will identify this area as a key monitoring area so that subsequent attention can be paid to the status changes of apples in this area.

[0063] In the embodiment of the present invention, the acquisition of monitoring area features in the food appearance image according to the candidate appearance detection features and the candidate appearance features may be implemented through the following examples.

[0064] Based on the candidate appearance detection features and the candidate appearance features, performing an optimization operation on the preset monitoring area features, and designating the optimized monitoring area features as the preset monitoring area features;

[0065] The step of extracting the appearance features of the target evaluation attribute from the appearance features based on the saliency evaluation result is repeated until all the appearance features are the candidate appearance features, thereby obtaining the monitoring area features.

[0066] In the embodiment of the present invention, for example, taking the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server has obtained the alternative appearance detection features and alternative appearance features of the apples, and now it is necessary to collect the monitoring area features based on these. The server has set some preset monitoring area features about apples in advance. These features may be preliminarily set based on the common perishable parts of apples or the areas where the appearance changes are prone to occur, such as some basic feature descriptions of the area near the apple stem and the area with slight bump marks on the skin. After obtaining the current alternative appearance detection features of the apple (for example, features that reflect the color change trend of a local area that conforms to possible deterioration) and alternative appearance features (such as the actual dull color and slightly irregular shape of the area), the server will conduct a comprehensive analysis of these features with the preset monitoring area features. Assume that the feature description of the area near the apple stem in the preset monitoring area features is relatively broad, and only roughly marks the range and some basic appearance conditions. Now, combining the alternative appearance detection features and the alternative appearance features, it is found that the color change of a certain part near the fruit stem is more obvious than expected and conforms to a certain deterioration trend (reflected by the alternative appearance detection features). At the same time, the actual appearance of the part also has some subtle texture changes (reflected by the alternative appearance features). The server will refine, supplement and adjust the description of the area near the fruit stem in the preset monitoring area features based on these new situations, such as more accurately determining the specific range of color changes, texture change details, etc. After completing the tuning operation, the tuned monitoring area feature of the area near the fruit stem is re-designated as a new preset monitoring area feature, so that further analysis and processing can be more accurately based on this in the future. The server returns to the appearance features of the apple again. It has previously made a saliency evaluation based on the evaluation attributes of the appearance features, and extracted the appearance features of some target evaluation attributes as alternative appearance features. However, there may be other parts of the appearance features that have not been fully considered. The server will analyze the remaining appearance features based on the evaluation attributes of the appearance features (such as shape, color, texture, etc.) according to the previous saliency evaluation process. For example, the first evaluation may have focused on areas with more obvious changes in color, while this time the focus may be on analyzing the prominence of the shape in the overall appearance. As this step is repeated over and over again, new appearance features with higher significance target evaluation attributes are extracted each time and added to the set of candidate appearance features. This process continues until all appearance features can be identified as candidate appearance features after evaluation, at which point the loop stops. Through this comprehensive and iterative evaluation and extraction process, the server eventually obtained monitoring area features that comprehensively consider all aspects of the apple's appearance features and more accurately reflect possible problem areas. These monitoring area features will provide an important basis for determining key monitoring areas.

[0067] In the embodiment of the present invention, the tuning operation is performed on the preset monitoring area feature based on the candidate appearance detection feature and the candidate appearance feature, which can be implemented through the following example.

[0068] Based on the candidate appearance detection features and the candidate appearance features, a tuning operation is performed on the preset monitoring area features to obtain original monitoring area features, and the original monitoring area features are designated as the preset monitoring area features;

[0069] The step of performing an optimization operation on the preset monitoring area feature based on the candidate appearance detection feature and the candidate appearance feature is looped until a preset cycle period is satisfied, thereby obtaining a monitoring area feature that has been optimized.

[0070] In the embodiment of the present invention, illustratively, still taking the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server has obtained the alternative appearance detection features, alternative appearance features and preset monitoring area features of the apples, and now the preset monitoring area features are to be tuned based on these. The preset monitoring area features initially set by the server are determined based on the general cognition of apples. For example, for apples, the area around the connection of the fruit stem and the area with slight bumps on the apple skin may be initially set as the preset monitoring area features. The description of these features is relatively broad and basic. When the current alternative appearance detection features of the apple are available (for example, through analysis, it is found that the color change trend of a local area of ​​the apple is consistent with the possible deterioration, and the texture also has slight changes, etc.) and the alternative appearance features (such as the actual color of the area is dull, the shape is slightly irregular, etc.), the server will use this information to tune the preset monitoring area features. Assume that the description of the area around the connection of the fruit stem in the original preset monitoring area features only roughly points out a range and briefly mentions some possible appearance changes. Now, combining the alternative appearance detection features and the alternative appearance features, it is found that the color change of a small area near the fruit stem connection is more obvious than previously expected, and the texture change is also clearer and more discernible (this information comes from the alternative appearance detection features and the alternative appearance features). Therefore, the server will refine and improve the original preset monitoring area features of the area around the fruit stem connection, such as more accurately determining the specific range of color changes, detailed characteristics of texture changes, etc. After such a tuning operation, a more accurate and detailed original monitoring area feature of the area around the fruit stem connection is obtained, and then this original monitoring area feature is re-designated as a new preset monitoring area feature, so that further tuning operations can be performed based on this in the future. The server sets a preset cycle period, for example, it stipulates that three such tuning cycle operations are to be performed. After completing the above first tuning operation and re-designating the preset monitoring area features, the next round of tuning operations will continue in the same way. As the apples are stored in the refrigerator for a longer time, their appearance features may change further, and the corresponding alternative appearance detection features and alternative appearance features may also be different. In each cycle, the server will again use the newly acquired candidate appearance detection features and candidate appearance features to tune the current preset monitoring area features (that is, the preset monitoring area features that were re-assigned after the previous round of tuning). For example, in the second cycle, it was found that another area of ​​the apple skin that had not been focused on before had new color changes and slight shape changes. The server will adjust the preset monitoring area features again based on these new situations. This cycle continues until the specified preset cycle period (such as three times) is completed. At this time, the monitoring area features that have been tuned multiple times and are more complete and accurate are obtained.These optimized monitoring area features can more accurately reflect the key areas on apples that may deteriorate, providing a more reliable basis for the subsequent determination of the true key monitoring areas.

[0071] In an embodiment of the present invention, the preset monitoring area feature includes at least one local feature of the monitoring area, and the tuning operation is performed on the preset monitoring area feature based on the alternative appearance detection feature and the alternative appearance feature to obtain the original monitoring area feature, which can be implemented through the following example.

[0072] Performing self-association strengthening and weighting on the local features of the monitoring area to obtain the features of the monitoring area to be processed;

[0073] Performing interactive correlation and strengthening weighting on the monitoring area features to be processed, the candidate appearance detection features and the candidate appearance features to obtain the candidate monitoring area features;

[0074] The candidate monitoring area features are aligned to the preset monitoring area feature vector dimensional space to obtain the original monitoring area features.

[0075] In the embodiment of the present invention, illustratively, continuing to take the apples in the refrigerator of the supermarket fresh food area as an example, the server has clarified the preset monitoring area features of the apple (including several local features of the monitoring area), the alternative appearance detection features and the alternative appearance features. Now, the tuning operation is performed based on these to obtain the original monitoring area features. Assume that the preset monitoring area features of the apple include the local features of the monitoring area around the fruit stem connection, such as the initial color of the apple skin in the area, the general shape characteristics and the basic description of the texture. The server will perform self-association reinforcement weighting on these local features of the monitoring area. Taking the skin color as an example, the server will analyze the degree of association between the colors at different positions in the area. If it is found that the color of a small area is slightly different from the surrounding area, and this difference may be of certain importance in judging the state of the apple, the color characteristics of this area will be given a higher weight to highlight its importance in the overall monitoring area local features. Similarly, for aspects such as shape characteristics and texture, corresponding weighting operations will be performed according to their own association in the area. After such self-correlation reinforcement weighting, the feature of the monitoring area to be processed is obtained, which highlights the factors in the local features of the monitoring area that may have an important impact on the judgment of the apple state. After obtaining the feature of the monitoring area to be processed, the server will perform cross-correlation reinforcement weighting together with the alternative appearance detection feature and the alternative appearance feature. For example, the color of the area around the connection of the fruit stalk in the feature of the monitoring area to be processed shows a certain trend of change, and the alternative appearance detection feature indicates that the area has color change characteristics that conform to a certain deterioration pattern, while the alternative appearance feature shows that the actual color of the area is dull and the shape is slightly irregular. The server will assign weights according to the relationship between them. If it is found that the color change trend in the feature of the monitoring area to be processed is highly consistent with the color change characteristics of the deterioration pattern in the alternative appearance detection feature, and is mutually confirmed with the actual dull color in the alternative appearance feature, the importance of these features in the cross-correlation will be given a higher weight. The same is true for shape, texture and other aspects. By comprehensively considering the interactive correlation between them and performing reasonable weighting operations on each feature, the alternative monitoring area feature is obtained. This feature more comprehensively integrates the correlation information between features from different sources and can more accurately reflect the actual situation of the apple in this area. The server previously set a specific vector dimension space for the preset monitoring area features to unify the specifications and process related features. Now we need to align the obtained alternative monitoring area features to this preset vector dimension space. For example, the preset monitoring area feature vector dimension space may specify a specific numerical range and representation method for describing the color feature, and there are corresponding specifications for the shape feature. The server will adjust the color, shape and other features in the alternative monitoring area features according to the requirements of this preset vector dimension space so that they can be represented under the same dimensional standard.After such an alignment operation, the original monitoring area features are obtained. On the basis of conforming to the preset vector dimension space specifications, it integrates more accurate feature information about the apple monitoring area obtained by previous operations, providing a more reliable basis for subsequent further processing and analysis.

[0076] In the embodiment of the present invention, the feature extraction process is performed on the food appearance image to obtain the appearance feature of at least one evaluation attribute, which can be implemented through the following examples.

[0077] Performing multi-gradient feature extraction processing on the food appearance image to obtain an original appearance feature of at least one target gradient;

[0078] Pre-processing the original appearance feature to obtain at least one to-be-processed appearance feature of the evaluation attribute;

[0079] According to a preset feature association strategy, feature cross processing is performed on the appearance features to be processed to obtain at least one appearance feature of an evaluation attribute.

[0080] In the embodiment of the present invention, for example, taking the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server needs to perform feature extraction processing on the acquired apple food appearance image, so as to obtain the appearance features that can be used for subsequent analysis and judgment. After receiving the food appearance image of the apple, the server starts to perform multi-gradient feature extraction processing. For example, for the image of the apple, the features will be analyzed and extracted from different levels of detail (i.e., gradients). At the low gradient level, it may focus on extracting relatively macroscopic features such as the overall shape of the apple (such as a circle) and the overall color distribution (such as red as the main color), which are the original appearance features under low gradients. At the medium gradient level, further attention will be paid to the texture trend of the apple skin, the depth change of local color, and other features, which are also used as the original appearance features of the corresponding gradient. At the high gradient level, it may be detailed to the location, size, and subtle differences in color of tiny spots on the apple skin, which are also extracted as the original appearance features of the high gradient. By analyzing and extracting the apple food appearance image from multiple gradients in this way, the server obtains a rich variety of original appearance features under different target gradients, and these original appearance features describe the appearance of the apple from different degrees of precision. After the server obtains the original appearance features of each target gradient, it needs to pre-process them so that they can be better classified and sorted according to the evaluation attributes. Taking the evaluation attribute of apple skin color as an example, the server will standardize the original appearance features involving color under different gradients extracted previously. For example, the color values ​​of different representations (which may come from different image analysis algorithms or devices) are unified into a standard representation form so that they can be accurately compared and analyzed later. At the same time, for the shape evaluation attribute, the description of the shape in the original appearance features will be optimized, such as refining some less accurate shape descriptions (such as roughly circular but somewhat fuzzy) into more accurate geometric shape descriptions (such as ellipse, with clear major and minor axis lengths, etc.). By performing similar pre-processing on the original appearance features corresponding to each evaluation attribute such as color and shape, the server obtains the to-be-processed appearance features of at least one evaluation attribute, and these to-be-processed appearance features have a more standardized and easier to process form in terms of their respective evaluation attributes. The server pre-sets a feature association strategy to guide how to further perform feature cross-processing on the to-be-processed appearance features that have been pre-processed. Assume that the preset feature association strategy stipulates that for the two evaluation attributes of the color and shape of the apple, the possible impact of the color feature on the shape feature should be analyzed based on the change of the color feature, and vice versa. According to this strategy, the server will associate the color change trend (such as the color gradually darkening) in the color to-be-processed appearance feature with the shape change possibility (such as the possibility of shrinking due to water loss) in the shape to-be-processed appearance feature.If it is found that the darker color area is often accompanied by a slight change in shape (such as becoming flatter), then the features of these two aspects are integrated to form a new appearance feature that combines color and shape information and belongs to a certain evaluation attribute (here it can be regarded as a newly defined comprehensive evaluation attribute). By performing such feature cross-processing on the to-be-processed appearance features of different evaluation attributes according to the preset feature association strategy, the server obtains the appearance features of at least one evaluation attribute. These appearance features more comprehensively reflect the relationship between the various aspects of the apple's appearance, providing a stronger basis for subsequent analysis and judgment.

[0081] In the embodiment of the present invention, the original can be implemented through the following examples.

[0082] Based on the target gradient, extracting the target original appearance feature from the original appearance feature, and optimizing the granularity of the target original appearance feature to obtain an optimized appearance feature;

[0083] Merging the optimized appearance feature with the preset spatial positioning feature to obtain a merged appearance feature, wherein the merged appearance feature includes at least one local image feature;

[0084] The local image features are weighted by self-association enhancement to obtain weighted appearance features, and based on the weighted appearance features, the original appearance features are tuned to obtain at least one to-be-processed appearance feature of the evaluation attribute.

[0085] In the embodiment of the present invention, illustratively, still taking the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server has obtained the original appearance features of the apples obtained through multi-gradient feature extraction processing, and now these original appearance features need to be pre-processed. After the server previously performed multi-gradient feature extraction processing on the appearance image of the apple food, it obtained the original appearance features of different target gradients, such as the overall shape of the apple, color distribution and other features under the low gradient, the direction of the skin texture, local color depth changes and other features under the medium gradient, and the location, size and color of the tiny spots on the skin. The high gradient has subtle differences. Features. Assuming that we are concerned about the original appearance features of the medium gradient, the server will extract the target original appearance features related to the current analysis focus based on this target gradient, such as the two features of the direction of the skin texture and the local color depth changes. Then, the granularity of the extracted target original appearance features is optimized. Taking the direction of the skin texture as an example, it may have only roughly described whether the texture was horizontal or vertical. Now, through more detailed analysis, it is refined into the specific angle of the texture, the degree of curvature, etc., and the local color depth changes are similarly refined, so that it can more accurately describe the actual appearance of the apple in this area, so as to obtain the optimized appearance features. The server stores preset spatial positioning features, which may be information about the placement and orientation of the apple in the refrigerator, as well as spatial reference information related to the image shooting angle. The optimized appearance features obtained before are merged with these preset spatial positioning features. For example, the optimized appearance features include the color depth changes and texture direction of a certain area of ​​the apple skin. The preset spatial positioning features indicate that the apple is placed horizontally in the middle layer of the refrigerator and the image is taken from the front. After merging, the merged appearance features are obtained, which not only contain detailed information about the appearance of the apple itself, but also incorporate relevant information such as its position in space and shooting angle, and at least one local image feature can be divided from it. For example, the appearance feature of a small area centered on the top of the apple is a local image feature. The server operates on each local image feature divided from the merged appearance features. Take the local image feature of the top of an apple as an example, which contains information such as the color and texture of the area. The server will analyze the degree of correlation between the elements within the local image feature, such as the relationship between color change and texture change. If it is found that the area with darker color is often accompanied by rougher texture, these closely related elements will be given higher weights, and their own correlation will be strengthened and weighted to obtain the weighted appearance features. Then, based on the obtained weighted appearance features, the server will perform tuning operations on the original appearance features.For example, if the weighted appearance features show that the top area of ​​the apple has become darker in color and rougher in texture, then the descriptions of the color, texture, etc. of this area in the original appearance features will be adjusted and optimized according to this situation, and finally at least one unprocessed appearance feature with an evaluation attribute will be obtained. These unprocessed appearance features, after combining multiple information such as spatial positioning and self-association, can more accurately reflect the appearance status of the apple and provide a more suitable basis for subsequent processing.

[0086] In an embodiment of the present invention, the preset feature association strategy includes a first feature association strategy and a second feature association strategy. According to the preset feature association strategy, feature cross-processing is performed on the appearance feature to be processed to obtain an appearance feature of at least one evaluation attribute, which can be implemented through the following example.

[0087] Based on the feature granularity of the appearance feature to be processed, performing a significance evaluation on the appearance feature to be processed to obtain a first granularity significance evaluation result;

[0088] According to the first granularity significance evaluation result, performing feature cross processing on the appearance feature to be processed based on the first feature association strategy to obtain at least one associated feature to be processed;

[0089] Based on the feature granularity of the associated feature to be processed, performing a significance evaluation on the associated feature to be processed to obtain a significance evaluation result of a second granularity;

[0090] According to the second granularity significance evaluation result, feature cross processing is performed on the to-be-processed associated features based on the second feature association strategy to obtain at least one associated feature, and each associated feature is designated as an appearance feature of an evaluation attribute.

[0091] In an embodiment of the present invention, illustratively, taking the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server has completed the pre-processing of the original appearance features of the apples and obtained the appearance features to be processed of at least one evaluation attribute. Now, these appearance features to be processed are subjected to feature cross-processing according to the preset feature association strategy. The server obtains the appearance features to be processed of the apples, such as the appearance features to be processed involving the evaluation attributes such as the color, texture, and shape of the apple skin. For the color appearance features to be processed, it may contain the specific numerical description of the color of different regions of the apple and the information such as the color change trend, which is its feature granularity. The server will analyze the prominence of these appearance features to be processed in the entire apple appearance based on the feature granularity. Taking the color appearance features to be processed as an example, if it is found that the color value of an area on the surface of the apple is greatly different from that of other areas, and the color change trend is obvious (such as from light red to dark red quickly), then the color appearance features to be processed in this area have a higher significance under the evaluation attribute of color. By performing similar analysis on the to-be-processed appearance features of various evaluation attributes such as color, texture, and shape, the server obtains the first granularity significance evaluation result, and clarifies the significance of each to-be-processed appearance feature under each evaluation attribute. Assume that the first feature association strategy stipulates that for the to-be-processed appearance features of the two evaluation attributes of color and shape, it is necessary to focus on the possible impact of color change on shape and vice versa. Based on the first granularity significance evaluation result, the server selects those to-be-processed appearance features with high significance under the color evaluation attribute, such as the to-be-processed appearance features corresponding to the areas with obvious color changes mentioned above. Then, combined with the to-be-processed appearance features of the shape evaluation attribute, the relationship between color change and shape change is analyzed. If it is found that the shape of the area with darker color is also slightly deflated, the features of these two aspects are integrated to obtain a to-be-processed associated feature, which integrates the relevant information of color and shape. By performing such feature cross-processing on the to-be-processed appearance features of different evaluation attributes according to the first feature association strategy, the server obtains at least one to-be-processed associated feature. The server obtains the unprocessed associated features obtained through the previous steps, such as the unprocessed associated features that have just integrated color and shape information. The feature granularity includes more comprehensive information such as color changes and shape changes. Also based on its feature granularity, the server analyzes the prominence of these unprocessed associated features in the entire apple appearance. If it is found that the color and shape changes of the apple area corresponding to this integrated unprocessed associated feature are more prominent in the entire apple appearance (for example, the changes are more obvious than other areas), then this unprocessed associated feature has a higher significance under the comprehensive evaluation attribute. By performing similar analysis on all the obtained unprocessed associated features, the server obtained the second granularity significance evaluation result, clarifying the significance of each unprocessed associated feature under the comprehensive evaluation attribute.Assume that the second feature association strategy stipulates that for the to-be-processed associated features that integrate color and shape information, further combined with the texture changes of the apple for comprehensive analysis. Based on the second granularity significance evaluation results, the server selects those to-be-processed associated features with high significance under the comprehensive evaluation attributes, such as the to-be-processed associated features with obvious and prominent color and shape changes mentioned above. Then, combined with the to-be-processed appearance features of the apple texture evaluation attribute, the relationship between color, shape changes and texture changes is analyzed. If it is found that the texture of the area where the color becomes darker and the shape becomes flat also becomes rough, the features of these three aspects are integrated to obtain an associated feature that integrates the relevant information of color, shape and texture. By performing such feature crossover on the to-be-processed associated features of different evaluation attributes according to the second feature association strategy, each associated feature is finally designated as an appearance feature of an evaluation attribute, such as designating this associated feature that integrates color, shape and texture information as a new evaluation attribute "appearance comprehensive change feature" so that these appearance features can be better used for analysis and judgment in the future.

[0092] In the embodiment of the present invention, according to the first granularity significance evaluation result, feature cross processing is performed on the appearance feature to be processed based on the first feature association strategy to obtain at least one associated feature to be processed, which can be implemented through the following examples.

[0093] Extracting the to-be-processed appearance feature with the smallest feature granularity from the to-be-processed appearance features to obtain candidate to-be-processed appearance features, and designating the candidate to-be-processed appearance features as the first to-be-processed associated features;

[0094] Based on the first granularity significance evaluation result, extracting the next to-be-processed appearance feature of the candidate to-be-processed appearance feature from the to-be-processed appearance feature to obtain a first to-be-associated feature of the candidate to-be-processed appearance feature;

[0095] Performing a feature integration operation on the candidate to-be-processed appearance feature and the first to-be-associated feature to obtain a second to-be-processed associated feature, and designating the first to-be-associated feature as the candidate to-be-processed appearance feature;

[0096] The step of extracting the next appearance feature to be processed of the candidate appearance feature to be processed from the appearance features to be processed based on the first granularity significance evaluation result is looped until all the appearance features to be processed complete feature intersection, and at least one associated feature to be processed is obtained, wherein the associated feature to be processed includes the first associated feature to be processed and the second associated feature to be processed.

[0097] In the embodiment of the present invention, the server obtains a series of appearance features to be processed about apples, which cover multiple evaluation attributes such as color, texture, and shape of the apple skin. For example, the color appearance feature to be processed may record in detail the color values ​​and color change trends of different areas of the apple; the texture appearance feature to be processed contains information such as the direction and density of the skin texture; and the shape appearance feature to be processed describes the overall and local shape characteristics of the apple. Among these appearance features to be processed, the server analyzes the feature granularity of each feature. Taking the color appearance feature to be processed as an example, if the color description of a certain area is simply represented by a rough color range (such as a red area), its feature granularity is smaller than other features that describe the color change value in more detail. The server compares the feature granularity of all appearance features to be processed, extracts the appearance feature to be processed with the smallest feature granularity, and assumes that it is the color appearance feature to be processed of a small area on the top of the apple (simply marked as a light red area), and uses it as an alternative appearance feature to be processed, and at the same time designates it as the first associated feature to be processed. The server already has the first-granularity significance evaluation result, which clarifies the significance of each to-be-processed appearance feature under its own evaluation attribute. Based on this evaluation result, the server searches for the next to-be-processed appearance feature associated with the candidate to-be-processed appearance feature (i.e., the color to-be-processed appearance feature of the light red area at the top of the apple just determined) among the remaining to-be-processed appearance features. Assuming that according to the first feature association strategy, for the color to-be-processed appearance feature, it is necessary to pay attention to its association with the shape to-be-processed appearance feature. Then the server will search for the part that may be associated with the light red area at the top of the apple in the shape to-be-processed appearance feature. For example, if the shape to-be-processed appearance feature at the corresponding position of the top of the apple shows that the shape of the area is slightly flat, this shape to-be-processed appearance feature is extracted as the first to-be-associated feature of the candidate to-be-processed appearance feature (the color to-be-processed appearance feature of the light red area at the top of the apple). The server now has the candidate to-be-processed appearance feature (the color to-be-processed appearance feature of the light red area at the top of the apple) and its first to-be-associated feature (the shape to-be-processed appearance feature of the slightly flat shape at the corresponding position of the top of the apple). Through a specific algorithm, the server integrates these two features. For example, the reddish area information in the color pending appearance feature and the slightly flat shape information in the shape pending appearance feature will be combined according to certain rules to form a new feature that combines the color and shape information. This feature is the second pending associated feature. At the same time, for the convenience of subsequent loop operations, the first pending associated feature (that is, the slightly flat shape pending appearance feature corresponding to the top of the apple) is designated as the new candidate pending appearance feature. The server will continue to loop according to the above steps. After completing the above integration operation, the new candidate pending appearance feature (the shape pending appearance feature just designated) becomes the starting point for the next round of loops.Based on the first granularity significance evaluation result, the server continues to search for the next to-be-processed appearance feature associated with the new candidate to-be-processed appearance feature among the remaining to-be-processed appearance features. Assume that this time, according to the first feature association strategy, for the shape to-be-processed appearance feature, it is necessary to pay attention to its association with the texture to-be-processed appearance feature. Then the server will look for a part in the texture to-be-processed appearance feature that may be associated with the slightly flat area in the shape corresponding to the top of the apple. For example, if it is found that the texture to-be-processed appearance feature in this area shows that the texture is relatively rough, this texture to-be-processed appearance feature is extracted as the first to-be-associated feature of the new candidate to-be-processed appearance feature (shape to-be-processed appearance feature in a slightly flat area). Then the integration operation is performed again to obtain a new second to-be-processed associated feature. This cycle continues, and the new candidate to-be-processed appearance feature and its associated features are integrated each time to continuously generate new to-be-processed associated features. The server stops the loop operation until all the appearance features to be processed have participated in such a feature crossover operation, that is, the feature crossover of all the appearance features to be processed is completed. At this time, the at least one associated feature to be processed obtained includes the first associated feature to be processed that was initially specified and the second associated features to be processed generated in the subsequent loop process. These associated features to be processed integrate the information of different evaluation attributes, providing a more comprehensive basis for subsequent further analysis and processing.

[0098] In the embodiment of the present invention, the step of performing a feature integration operation on the candidate appearance feature to be processed and the feature to be associated to obtain a second associated feature to be processed may be implemented through the following example.

[0099] Standardizing the feature attributes of the candidate to-be-processed appearance features to obtain standardized appearance features of preset attribute dimensions;

[0100] Performing a feature granularity enhancement process on the standardized appearance feature to obtain an enhanced appearance feature, wherein the feature granularity of the enhanced appearance feature is the same as the feature granularity of the first feature to be associated;

[0101] Standardizing the first feature to be associated with a feature attribute to obtain the first feature to be associated with a standardization of the preset attribute dimension;

[0102] The improved appearance feature is combined with the standardized first feature to be associated to obtain a second associated feature to be processed.

[0103] In the embodiment of the present invention, for example, the apples in the refrigerator in the fresh food area of ​​the supermarket are taken as an example. The server has determined the candidate appearance feature to be processed and the corresponding first feature to be associated in the previous operation. Now, the server needs to perform a feature integration operation on them to obtain the second associated feature to be processed. Assuming that the selected appearance feature to be processed is the color appearance feature to be processed of a certain area on the top of the apple, the original description of the color may be in a relatively simple and less standardized way, such as just using a relatively broad expression such as "light red" to represent the color of the area, and it may not be uniform in terms of data format or dimension. The server will perform feature attribute standardization on the color appearance feature to be processed. It will convert the color description into a more accurate and standardized form according to the pre-set standard, such as using a specific color value (such as RGB value) to accurately represent the color of the area, and make its data format and dimension meet the requirements of the preset attribute dimension. After such processing, the standardized appearance feature of the preset attribute dimension is obtained. At this time, the color feature of the area is more accurate and standardized in data representation, which is convenient for subsequent integration operations with other features. After the previous standardization process, the obtained standardized appearance features are more standardized in data representation, but their feature granularity may still be relatively fine, which is not very compatible with the feature granularity of the first feature to be associated to be integrated next. For example, the standardized appearance features may only accurately describe the color itself, while the first feature to be associated (assuming that it is the shape of the area to be processed appearance feature, which describes the slightly flat shape of the area) contains more detailed information about the shape of the area in terms of feature granularity, such as the specific aspect ratio of the shape, the degree of curvature of the edge, etc. The server will improve the feature granularity of the standardized appearance features. For the color to be processed appearance feature, it may further analyze the contrast between the color of the area and the color of the surrounding area, the uniformity of the color distribution in the area, etc., so as to add more relevant detailed information on the basis of the original description of the color itself, so that its feature granularity is improved to a level similar to that of the first feature to be associated, and the appearance feature to be improved is obtained. For the first feature to be associated, that is, the shape to be processed appearance feature of the area on the top of the apple, its original description of the shape may also have some irregularities or inconsistent data formats. For example, a description of a slightly flat shape may be just a more intuitive description, not represented by precise geometric data. The server will perform feature attribute standardization on the appearance feature to be processed of this shape. According to the requirements of the preset attribute dimension, the description of the shape is converted into precise geometric data (such as the length of the major axis and minor axis of the ellipse), and the data format is made to meet the standard, so as to obtain the first feature to be associated after standardization of the preset attribute dimension. In this way, the shape feature is more accurate and standardized in data representation, which is convenient for subsequent integration operations.The server now has the appearance features that have been improved (the color appearance features to be processed after granularity improvement) and the first features to be associated after standardization (the shape appearance features to be processed after standardization). The server merges these two features through a specific algorithm. For example, the detailed information about color in the appearance features that have been improved (such as color value, distribution, etc.) and the detailed information about shape in the first features to be associated after standardization (such as shape geometry data, edge conditions, etc.) will be combined according to certain rules to form a new feature that combines color and shape information. This feature is the second associated feature to be processed, which provides a more comprehensive basis for subsequent further analysis and processing.

[0104] In the embodiment of the present invention, according to the second granularity significance evaluation result, feature cross processing is performed on the to-be-processed associated features based on the second feature association strategy to obtain at least one associated feature, which can be implemented through the following example.

[0105] Extracting the to-be-processed associated features with the largest feature granularity from the to-be-processed associated features to obtain candidate associated features, and performing a feature integration operation on features with different feature attributes in the candidate associated features to obtain a first associated feature;

[0106] Based on the second granularity significance evaluation result, extracting the next to-be-processed associated feature of the candidate associated feature from the to-be-processed associated features to obtain a second to-be-processed associated feature of the candidate associated feature;

[0107] Performing a feature integration operation on the candidate association feature and the second feature to be associated to obtain a second association feature, and designating the second feature to be associated as the candidate association feature;

[0108] The step of extracting the next associated feature to be processed of the candidate associated feature from the associated features to be processed based on the second granularity significance evaluation result is looped until all associated features to be processed complete feature intersection, and at least one associated feature is obtained, wherein the associated feature includes the first associated feature and the second associated feature.

[0109] In the embodiment of the present invention, illustratively, still taking the apples in the refrigerator in the fresh food area of ​​the supermarket as an example, the server has completed the second granularity significance evaluation of the to-be-processed associated features, and now it is necessary to perform feature cross-processing on these to-be-processed associated features according to the second feature association strategy to obtain the associated features. The server obtains a series of to-be-processed associated features about apples, which are previously obtained by processing based on the first feature association strategy, and integrate information of different evaluation attributes such as color and shape. The server analyzes the feature granularity of each to-be-processed associated feature. For example, a to-be-processed associated feature records in detail the color change of a certain area of ​​the apple (including specific color values, change trends, etc.), shape change (such as a specific data description of a change from a round shape to a slightly flat shape), and texture change (such as a description of the degree of texture from a fine shape to a rough shape). Its feature granularity is richer than other to-be-processed associated features and contains more detailed information. The server extracts the to-be-processed associated feature with the largest feature granularity from these to-be-processed associated features as an alternative associated feature. Then, for this alternative associated feature, it contains features of different feature attributes (such as color, shape, and texture). The server will perform feature integration operations on these features with different feature attributes according to a specific algorithm. For example, the relevant information of color change, shape change and texture change is combined according to certain rules to form a new feature that integrates color, shape and texture information. This feature is the first associated feature. The server is based on the completed second granularity significance evaluation result, which clarifies the significance of each associated feature to be processed under the comprehensive evaluation attribute. Based on this result, the next associated feature to be processed that is associated with the candidate associated feature is found in the remaining associated features to be processed. Assume that according to the second feature association strategy, for the alternative associated feature that just obtained that integrates color, shape and texture information, it is necessary to pay attention to its association with the relevant features of the apple stalk. Then the server will look for parts that may be associated with the apple stalk in other associated features to be processed. For example, if it is found that a to-be-processed associated feature describes that the color change and shape change at the stalk are related to the area involved in the alternative associated feature just now, this to-be-processed associated feature is extracted as the second to-be-associated feature of the alternative associated feature. The server now has a candidate association feature (the feature that just combined the color, shape and texture information) and its second feature to be associated (the pending association feature related to the fruit stem). These two features are integrated through a specific algorithm. For example, the color, shape, and texture information in the candidate association feature and the fruit stem related information in the second feature to be associated (such as color changes, shape changes, etc.) will be combined according to certain rules to form a new feature that combines more information. This feature is the second association feature. At the same time, in order to facilitate subsequent cyclic operations, the second feature to be associated is designated as the new candidate association feature.The server will continue to loop according to the above steps. After completing the above integration operation, the new candidate association feature (the pending association feature just specified and related to the fruit stalk) becomes the starting point of the next cycle. Based on the second granularity significance evaluation result, the server continues to search for the next pending association feature associated with the new candidate association feature among the remaining pending association features. Then the integration operation is performed again to obtain a new second association feature. This cycle continues, and the new candidate association feature and its associated features are integrated each time to continuously generate new association features. Until all pending association features participate in such a feature crossover operation, that is, when the feature crossover of all pending association features is completed, the server stops the loop operation. At this time, at least one association feature obtained includes the first association feature initially obtained and the second association features generated in the subsequent cycle process. These association features integrate more information on different evaluation attributes, providing a more comprehensive basis for subsequent further analysis and processing.

[0110] Please refer to Figure 2 , Figure 2 The structure schematic block diagram of the intelligent fresh-keeping system 110 for food refrigeration equipment provided in an embodiment of the present invention includes:

[0111] The acquisition module 1101 is used to acquire an image of the target food through an image sensor of the food refrigeration equipment to obtain an initial food image;

[0112] The recognition module 1102 is used to input the initial food image into a pre-trained food category recognition model for processing, to obtain the appearance and food type of the target food to be monitored; based on the appearance of the food to be monitored, to determine at least one key monitoring area;

[0113] The fresh-keeping module 1103 is used to collect image data of at least one key monitoring area, and match it in a preset fresh-keeping strategy library in combination with the food type to obtain the current fresh-keeping strategy for the target food; and send the current fresh-keeping strategy to the central control chip of the food refrigeration equipment to execute the current fresh-keeping strategy.

[0114] It should be noted that the implementation principle of the intelligent fresh-keeping system 110 of the aforementioned food refrigeration equipment can refer to the implementation principle of the intelligent fresh-keeping method of the aforementioned food refrigeration equipment, which will not be repeated here. It should be understood that the division of the various modules of the above device is only a division of logical functions, and in actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated.

[0115] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned intelligent fresh-keeping system 110 of the food refrigeration equipment. Figure 3 As shown, Figure 3 The computer device 100 provided in the embodiment of the present invention is a structural block diagram. The computer device 100 includes an intelligent fresh-keeping system 110 for food refrigeration equipment, a memory 111, a processor 112 and a communication unit 113.

[0116] In order to realize data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are electrically connected to each other directly or indirectly. For example, the electrical connection between these elements can be realized through one or more communication buses or signal lines. The intelligent fresh-keeping system 110 of the food refrigeration equipment includes at least one software function module that can be stored in the memory 111 in the form of software or firmware or solidified in the operating system (OS) of the computer device 100. The processor 112 is used to execute the intelligent fresh-keeping system 110 of the food refrigeration equipment stored in the memory 111, such as the software function modules and computer programs included in the intelligent fresh-keeping system 110 of the food refrigeration equipment.

[0117] For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. Numerous modifications and variations are possible in accordance with the above teachings. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. The intelligent fresh-keeping system of food refrigeration equipment is characterized by: include: An acquisition module is used to acquire an image of a target food through an image sensor of a food refrigeration device to obtain an initial food image; A recognition module, used to input the initial food image into a pre-trained food category recognition model for processing, to obtain the appearance and food type of the target food to be monitored; based on the appearance of the food to be monitored, to determine at least one key monitoring area; A fresh-keeping module, used to collect image data of the at least one key monitoring area, and match the food type in a preset fresh-keeping strategy library to obtain a current fresh-keeping strategy for the target food; and send the current fresh-keeping strategy to the central control chip of the food refrigeration equipment to execute the current fresh-keeping strategy; The identification module is specifically used for: Acquire the appearance of a food to be monitored and food deterioration description information corresponding to the appearance of the food to be monitored, wherein the appearance of the food to be monitored includes at least one food appearance image; Performing feature extraction processing on the food appearance image to obtain at least one appearance feature of the evaluation attribute, and performing feature extraction processing on the food deterioration description information to obtain monitoring content features; the monitoring content features include at least one monitoring indicator feature; Identifying feature attribute dimensions in the appearance features to obtain appearance attribute dimensions, and identifying feature attribute dimensions in the monitored content features to obtain content attribute dimensions; Based on the appearance attribute dimension and the content attribute dimension, determining the target attribute dimension of the vector dimension space corresponding to each evaluation attribute, and performing a feature integration operation on the monitoring indicator feature to obtain an integrated monitoring content feature; Align the attribute dimensions of the appearance feature and the attribute dimensions of the integrated monitoring content feature to the target attribute dimension, respectively, to obtain the target appearance feature of each evaluation attribute and the target monitoring content feature corresponding to the target appearance feature; Collecting local image features corresponding to each image grid unit in the target appearance features; Collecting features under each feature attribute in the local image features to obtain appearance attribute features; Extracting features under the feature attributes corresponding to the appearance attribute features from the target monitoring content features to obtain content attribute features; Performing a feature integration operation on the appearance attribute feature and the content attribute feature to obtain an integrated attribute feature corresponding to each feature attribute; Superimposing the integrated attribute features under each image grid unit to obtain a first integrated feature corresponding to each image grid unit; Performing a feature integration operation on the local image feature and the target monitoring content feature to obtain a second integrated feature, and determining a relative ratio between the second integrated feature and the first integrated feature to obtain a feature matching coefficient corresponding to each image grid unit, wherein the feature matching coefficient is used to reflect the matching situation between the appearance feature and the monitoring content feature in the same image grid unit; Based on the feature matching coefficient, the local image features are weighted to obtain original appearance detection features corresponding to each image grid unit; Performing a feature integration operation on the original appearance detection features of the same evaluation attribute to obtain appearance detection features of each evaluation attribute; Based on the evaluation attribute of the appearance feature, performing a saliency evaluation on the appearance feature, and extracting the appearance feature of the target evaluation attribute from the appearance feature based on the saliency evaluation result to obtain a candidate appearance feature; Extracting the appearance detection features corresponding to the target evaluation attributes from the appearance detection features to obtain candidate appearance detection features; Based on the candidate appearance detection features and the candidate appearance features, performing an optimization operation on the preset monitoring area features, and designating the optimized monitoring area features as the preset monitoring area features; The step of extracting the appearance features of the target evaluation attribute from the appearance features based on the significance evaluation result is repeated until all the appearance features are the candidate appearance features, thereby obtaining monitoring area features, and determining at least one key monitoring area in the food appearance image based on the monitoring area features.

2. The system according to claim 1, characterized in that The identification module is further specifically used for: Based on the candidate appearance detection features and the candidate appearance features, a tuning operation is performed on the preset monitoring area features to obtain original monitoring area features, and the original monitoring area features are designated as the preset monitoring area features; The step of performing an optimization operation on the preset monitoring area feature based on the candidate appearance detection feature and the candidate appearance feature is looped until a preset cycle period is satisfied, thereby obtaining a monitoring area feature that has been optimized.

3. The system according to claim 2, characterized in that The preset monitoring area feature includes at least one local feature of the monitoring area, and the identification module is further specifically used for: Performing self-association strengthening and weighting on the local features of the monitoring area to obtain the features of the monitoring area to be processed; Performing interactive correlation and strengthening weighting on the monitoring area features to be processed, the candidate appearance detection features and the candidate appearance features to obtain the candidate monitoring area features; The candidate monitoring area features are aligned to the preset monitoring area feature vector dimensional space to obtain the original monitoring area features.

4. The system according to claim 1, characterized in that The identification module is further specifically used for: Performing multi-gradient feature extraction processing on the food appearance image to obtain an original appearance feature of at least one target gradient; Based on the target gradient, extracting the target original appearance feature from the original appearance feature, and optimizing the granularity of the target original appearance feature to obtain an optimized appearance feature; Merging the optimized appearance feature with the preset spatial positioning feature to obtain a merged appearance feature, wherein the merged appearance feature includes at least one local image feature; Performing self-association enhancement weighting on the local image features to obtain weighted appearance features, and performing a tuning operation on the original appearance features based on the weighted appearance features to obtain a to-be-processed appearance feature of at least one evaluation attribute; According to a preset feature association strategy, feature cross processing is performed on the appearance features to be processed to obtain at least one appearance feature of an evaluation attribute.

5. The system according to claim 4, characterized in that The preset feature association strategy includes a first feature association strategy and a second feature association strategy, and the identification module is further specifically used for: Based on the feature granularity of the appearance feature to be processed, performing a significance evaluation on the appearance feature to be processed to obtain a first granularity significance evaluation result; According to the first granularity significance evaluation result, performing feature cross processing on the appearance feature to be processed based on the first feature association strategy to obtain at least one associated feature to be processed; Based on the feature granularity of the associated feature to be processed, performing a significance evaluation on the associated feature to be processed to obtain a significance evaluation result of a second granularity; According to the second granularity significance evaluation result, feature cross processing is performed on the to-be-processed associated features based on the second feature association strategy to obtain at least one associated feature, and each associated feature is designated as an appearance feature of an evaluation attribute.

6. The system according to claim 5, characterized in that The identification module is further specifically used for: Extracting the to-be-processed appearance feature with the smallest feature granularity from the to-be-processed appearance features to obtain candidate to-be-processed appearance features, and designating the candidate to-be-processed appearance features as the first to-be-processed associated features; Based on the first granularity significance evaluation result, extracting the next to-be-processed appearance feature of the candidate to-be-processed appearance feature from the to-be-processed appearance feature to obtain a first to-be-associated feature of the candidate to-be-processed appearance feature; Performing a feature integration operation on the candidate to-be-processed appearance feature and the first to-be-associated feature to obtain a second to-be-processed associated feature, and designating the first to-be-associated feature as the candidate to-be-processed appearance feature; The step of extracting the next appearance feature to be processed of the candidate appearance feature to be processed from the appearance features to be processed based on the first granularity significance evaluation result is looped until all the appearance features to be processed complete feature intersection, and at least one associated feature to be processed is obtained, wherein the associated feature to be processed includes the first associated feature to be processed and the second associated feature to be processed.

7. The system according to claim 6, characterized in that The identification module is further specifically used for: Standardizing the feature attributes of the candidate to-be-processed appearance features to obtain standardized appearance features of preset attribute dimensions; Performing a feature granularity enhancement process on the standardized appearance feature to obtain an enhanced appearance feature, wherein the feature granularity of the enhanced appearance feature is the same as the feature granularity of the first feature to be associated; Standardizing the first feature to be associated with a feature attribute to obtain the first feature to be associated with a standardization of the preset attribute dimension; The improved appearance feature is combined with the standardized first feature to be associated to obtain a second associated feature to be processed.

8. The system according to claim 5, characterized in that The identification module is further specifically used for: Extracting the to-be-processed associated features with the largest feature granularity from the to-be-processed associated features to obtain candidate associated features, and performing a feature integration operation on features with different feature attributes in the candidate associated features to obtain a first associated feature; Based on the second granularity significance evaluation result, extracting the next to-be-processed associated feature of the candidate associated feature from the to-be-processed associated features to obtain a second to-be-processed associated feature of the candidate associated feature; Performing a feature integration operation on the candidate association feature and the second feature to be associated to obtain a second association feature, and designating the second feature to be associated as the candidate association feature; The step of extracting the next to-be-processed associated feature of the candidate associated feature from the to-be-processed associated features based on the second granularity significance evaluation result is looped until all to-be-processed associated features complete feature intersection, and at least one associated feature is obtained, wherein the associated features include the first associated feature and the second associated feature.

9. An intelligent fresh-keeping method for food refrigeration equipment, characterized in that: include: Capturing an image of the target food through an image sensor of the food refrigeration equipment to obtain an initial food image; Inputting the initial food image into a pre-trained food category recognition model for processing to obtain the appearance and food type of the target food to be monitored; Based on the appearance of the food to be monitored, determining at least one key monitoring area; Collecting image data of the at least one key monitoring area, and matching the image data in a preset fresh-keeping strategy library in combination with the food type, to obtain a current fresh-keeping strategy for the target food; Sending the current fresh-keeping strategy to the central control chip of the food refrigeration equipment to execute the current fresh-keeping strategy; Determining at least one key monitoring area based on the appearance of the food to be monitored includes: Acquire the appearance of a food to be monitored and food deterioration description information corresponding to the appearance of the food to be monitored, wherein the appearance of the food to be monitored includes at least one food appearance image; Performing feature extraction processing on the food appearance image to obtain at least one appearance feature of the evaluation attribute, and performing feature extraction processing on the food deterioration description information to obtain monitoring content features; the monitoring content features include at least one monitoring indicator feature; Identifying feature attribute dimensions in the appearance features to obtain appearance attribute dimensions, and identifying feature attribute dimensions in the monitored content features to obtain content attribute dimensions; Based on the appearance attribute dimension and the content attribute dimension, determining the target attribute dimension of the vector dimension space corresponding to each evaluation attribute, and performing a feature integration operation on the monitoring indicator feature to obtain an integrated monitoring content feature; Align the attribute dimensions of the appearance feature and the attribute dimensions of the integrated monitoring content feature to the target attribute dimension, respectively, to obtain the target appearance feature of each evaluation attribute and the target monitoring content feature corresponding to the target appearance feature; Collecting local image features corresponding to each image grid unit in the target appearance features; Collecting features under each feature attribute in the local image features to obtain appearance attribute features; Extracting features under the feature attributes corresponding to the appearance attribute features from the target monitoring content features to obtain content attribute features; Performing a feature integration operation on the appearance attribute feature and the content attribute feature to obtain an integrated attribute feature corresponding to each feature attribute; Superimposing the integrated attribute features under each image grid unit to obtain a first integrated feature corresponding to each image grid unit; Performing a feature integration operation on the local image feature and the target monitoring content feature to obtain a second integrated feature, and determining a relative ratio between the second integrated feature and the first integrated feature to obtain a feature matching coefficient corresponding to each image grid unit, wherein the feature matching coefficient is used to reflect the matching situation between the appearance feature and the monitoring content feature in the same image grid unit; Based on the feature matching coefficient, the local image features are weighted to obtain original appearance detection features corresponding to each image grid unit; Performing a feature integration operation on the original appearance detection features of the same evaluation attribute to obtain appearance detection features of each evaluation attribute; Based on the evaluation attribute of the appearance feature, performing a saliency evaluation on the appearance feature, and extracting the appearance feature of the target evaluation attribute from the appearance feature based on the saliency evaluation result to obtain a candidate appearance feature; Extracting the appearance detection features corresponding to the target evaluation attributes from the appearance detection features to obtain candidate appearance detection features; Based on the candidate appearance detection features and the candidate appearance features, performing an optimization operation on the preset monitoring area features, and designating the optimized monitoring area features as the preset monitoring area features; The step of extracting the appearance features of the target evaluation attribute from the appearance features based on the significance evaluation result is repeated until all the appearance features are the candidate appearance features, thereby obtaining monitoring area features, and determining at least one key monitoring area in the food appearance image based on the monitoring area features.

Images