Preservation parameter adaptive optimization control system based on fruit state detection

Abstract

The invention relates to the technical field of biological preservation, and particularly discloses a preservation parameter adaptive optimization control system based on fruit state detection, which utilizes a camera to carry out state monitoring on a fruit object, and adopts an artificial intelligence technology based on deep learning to carry out image analysis on a surface state image of the fruit object. According to the method, the multi-scale feature representation of the fruit surface state is mined, intelligent identification of the type label of the fruit surface state is further carried out, and the appropriate oxygen concentration value in the current state is determined in combination with a fruit state label-recommended oxygen concentration corresponding table obtained through an experiment, so that the self-adaptive optimization control of the oxygen concentration is realized. Therefore, the preservation effect can be effectively improved, the shelf life of the fruits is prolonged, and the loss caused by improper preservation is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biological preservation, and more particularly, to a preservation parameter adaptive optimization control system based on fruit state detection. BACKGROUND

[0002] In the modern fruit and vegetable industry, preservation technology is one of the key factors to promote the quality and market competitiveness of agricultural products. With consumers paying more and more attention to the freshness, taste and nutritional value of food, how to effectively extend the shelf life of fruits and reduce losses has become an important problem to be solved in the industry.

[0003] Traditional fruit preservation methods mainly rely on manual experience judgment and fixed parameter setting, and the storage conditions are relatively fixed, lacking dynamic response capability to fruit state and optimization capability to storage conditions. This leads to suboptimal preservation results in actual application, which cannot meet the individualized preservation needs of different types and different maturity fruits.

[0004] Therefore, a preservation parameter adaptive optimization control system based on fruit state detection is expected. SUMMARY

[0005] To solve the above technical problems, the present application is proposed. The embodiments of the present application provide a preservation parameter adaptive optimization control system based on fruit state detection, which uses a camera to monitor the state of a fruit object, and uses artificial intelligence technology based on deep learning to analyze the surface state image of the fruit object to extract multi-scale feature representation of the fruit surface state, and then intelligently identifies the fruit surface state category label, and determines the appropriate oxygen concentration value under the current state by combining the fruit state label-recommended oxygen concentration correspondence table obtained through experiments, thereby realizing adaptive optimization control of oxygen concentration. In this way, the preservation effect can be effectively improved, the shelf life of fruits can be extended, and losses caused by improper preservation can be reduced.

[0006] Correspondingly, according to one aspect of the present application, a preservation parameter adaptive optimization control system based on fruit state detection is provided, which comprises: a specific fruit object state monitoring module for acquiring a surface state image of a specific fruit object through a camera; an image transmission module for inputting the surface state image of the specific fruit object into a parameter optimization control center, wherein the parameter optimization control center stores a fruit state label-recommended oxygen concentration correspondence table; a fruit state multi-scale feature extraction module for performing multi-scale feature extraction on the surface state image of the specific fruit object in the parameter optimization control center to obtain a fruit state shallow feature map and a fruit state semantic feature map; A fruit state shallow feature salient module is used to perform shallow feature salient on the fruit state shallow feature map in the parameter optimization control center to obtain a fruit state shallow salient feature map; A fruit state recognition module is used to generate a fruit state recognition result based on the multi-scale fusion features between the fruit state shallow salient feature map and the fruit state semantic feature map in the parameter optimization control center; The oxygen concentration adaptive adjustment module is used to generate an oxygen concentration adaptive adjustment instruction in the parameter optimization control center based on the recognition result and the fruit status label-recommended oxygen concentration correspondence table.

[0007] Compared with the existing technology, the adaptive optimization control system for preservation parameters based on fruit state detection provided by this application uses a camera to monitor the state of the fruit object and adopts deep learning-based artificial intelligence technology to perform image analysis on the surface state image of the fruit object to mine the multi-scale feature representation of the fruit surface state, and then use this to perform intelligent recognition of the fruit surface state category label. At the same time, combined with the fruit state label-recommended oxygen concentration correspondence table obtained through experiments, the appropriate oxygen concentration value under the current state is determined, thereby realizing adaptive optimization control of oxygen concentration. In this way, the preservation effect can be effectively improved, the shelf life of the fruit can be extended, and the losses caused by improper preservation can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a block diagram of a fresh-keeping parameter adaptive optimization control system based on fruit status detection according to an embodiment of the present application.

[0010] Figure 2 Schematic diagram of data flow of a fresh-keeping parameter adaptive optimization control system based on fruit status detection according to an embodiment of the present application.

[0011] Figure 3 This is a block diagram of a fruit state multi-scale feature extraction module in a fresh-keeping parameter adaptive optimization control system based on fruit state detection according to an embodiment of the present application.

[0012] Figure 4 This is a block diagram of a fruit state recognition module in a fresh-keeping parameter adaptive optimization control system based on fruit state detection according to an embodiment of the present application. DETAILED DESCRIPTION

[0013] Below, the embodiments of the present application will be described in more detail with reference to the accompanying drawings, and the above-mentioned and other purposes, features, and advantages of the present application will become more apparent. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described herein.

[0014] In response to the limitations of the existing technology described in the above background technology, the technical concept of this application is to use a camera to monitor the status of fruit objects, and use deep learning-based artificial intelligence technology to perform image analysis on the surface status images of fruit objects to extract multi-scale feature representations of the fruit surface status, and then use this to perform intelligent recognition of the fruit surface status category label. At the same time, combined with the fruit status label-recommended oxygen concentration correspondence table obtained through experiments, the appropriate oxygen concentration value under the current state is determined, thereby achieving adaptive optimization control of oxygen concentration. In this way, the preservation effect can be effectively improved, the shelf life of fruits can be extended, and the losses caused by improper preservation can be reduced.

[0015] Figure 1 This is a block diagram of a fresh-keeping parameter adaptive optimization control system based on fruit status detection according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the adaptive optimization control system for fresh-keeping parameters based on fruit status detection according to an embodiment of the present application. Figure 1 and Figure 2 As shown, according to the embodiment of the present application, the adaptive optimization control system 100 for preserving parameters based on fruit state detection includes: a specific fruit object state monitoring module 110, which is used to collect the surface state image of the specific fruit object through a camera; an image transmission module 120, which is used to input the surface state image of the specific fruit object into the parameter optimization control center, wherein the parameter optimization control center stores a fruit state label-recommended oxygen concentration correspondence table; a fruit state multi-scale feature extraction module 130, which is used to perform multi-scale feature extraction on the surface state image of the specific fruit object in the parameter optimization control center to obtain a shallow feature map of the fruit state and a fruit state shallow feature map. a fruit state semantic feature map; a fruit state shallow feature saliency module 140, for performing shallow feature saliency on the fruit state shallow feature map in the parameter optimization control center to obtain a fruit state shallow salient feature map; a fruit state recognition module 150, for generating a fruit state recognition result based on the multi-scale fusion features between the fruit state shallow salient feature map and the fruit state semantic feature map in the parameter optimization control center; an oxygen concentration adaptive adjustment module 160, for generating an oxygen concentration adaptive adjustment instruction based on the recognition result and the fruit state label-recommended oxygen concentration correspondence table in the parameter optimization control center.

[0016] In the above-mentioned fresh-keeping parameter adaptive optimization control system 100 based on fruit state detection, the specific fruit object state monitoring module 110 is configured to capture the surface state image of the specific fruit object by using a camera. It should be understood that the camera can capture the changes of the fruit surface in real time and obtain various information of the fruit surface, such as color, texture, spots, and degree of decay, so as to understand the state changes and fresh-keeping needs of the fruit in real time. Moreover, as a non-contact sensor, the camera can obtain the surface state information of the fruit without directly contacting the fruit, thereby effectively maintaining the integrity of the fruit and avoiding cross contamination.

[0017] In the above-mentioned fresh-keeping parameter adaptive optimization control system 100 based on fruit state detection, the image transmission module 120 is configured to input the surface state image of the specific fruit object into a parameter optimization control center, wherein the parameter optimization control center stores a fruit state label-recommended oxygen concentration correspondence table. It should be understood that the fruit state label-recommended oxygen concentration correspondence table is established based on previous experimental data, and contains the recommended oxygen concentration values under different fruit states, thereby providing a scientific basis for the adaptive adjustment of the oxygen concentration. That is, by matching the current fruit state with the correspondence table, the required oxygen concentration under the current fruit state can be quickly and accurately determined.

[0018] In the above-mentioned fresh-keeping parameter adaptive optimization control system 100 based on fruit state detection, the fruit state multi-scale feature extraction module 130 is configured to perform multi-scale feature extraction on the surface state image of the specific fruit object in the parameter optimization control center to obtain a fruit state shallow feature map and a fruit state semantic feature map. Wherein, Figure 3 A block diagram of the fruit state multi-scale feature extraction module in the fresh-keeping parameter adaptive optimization control system based on fruit state detection according to the embodiment of the present application is shown in FIG. 13. As shown in FIG. 13, the fruit state multi-scale feature extraction module 130 includes an image noise reduction unit 131 configured to perform noise reduction processing on the surface state image of the specific fruit object to obtain a noise-reduced fruit surface state image; and a multi-scale feature extraction unit 132 configured to input the noise-reduced fruit surface state image into a fruit surface state multi-scale feature extractor based on a dilated pyramid network to obtain the fruit state shallow feature map and the fruit state semantic feature map. Figure 3

[0019] ​Specifically, the image denoising unit 131 is configured to perform denoising processing on the surface state image of the specific fruit object to obtain a denoised fruit surface state image. It should be understood that the surface state image of the specific fruit object captured by the camera can contain noise interference, such as image blur or distortion caused by factors such as light variation and device jitter. Therefore, in order to improve the image quality, the surface state image of the specific fruit object is further subjected to denoising processing, so that the image can more accurately reflect the real state of the fruit surface, and provide more reliable input data for fruit state label recognition. In the embodiments of the present application, methods such as Gaussian filtering, median filtering or wavelet transform can be used to perform denoising processing on the surface state image of the specific fruit object to obtain a denoised fruit surface state image.

[0020] In the technical solution of the present application, the image denoising unit 131 is further configured to perform guided filter-based preliminary denoising processing on the surface state image of the specific fruit object to obtain a preliminary denoised image, and to perform adaptive weighted fusion based on noise evaluation on the preliminary denoised image and the surface state image of the specific fruit object to obtain the denoised fruit surface state image. Specifically, the image denoising unit is further configured to calculate a noise residual image between the preliminary denoised image and the surface state image of the specific fruit object, to perform normalization processing on the noise residual image to obtain an adaptive weight image, and to perform weighted fusion on the preliminary denoised image and the surface state image of the specific fruit object based on the adaptive weight image to obtain the denoised fruit surface state image.

[0021] More specifically, the image denoising unit is further configured to perform weighted fusion on the preliminary denoised image and the surface state image of the specific fruit object to obtain the denoised fruit surface state image according to the following formula: wherein, is the weight value of each position in the adaptive weight image, is the pixel value of each position in the surface state image of the specific fruit object, is the pixel value of each position in the preliminary denoised image.

[0022] Correspondingly, in the technical solution of the present application, in the adaptive optimization control process of fruit preservation parameters, the first step is to analyze the surface state image of a specific fruit object. However, the original surface state image obtained by the image acquisition device such as a camera will inevitably be affected by factors such as sensor thermal noise, uneven lighting, and environmental interference, thereby introducing various mixed noises such as Gaussian noise and salt and pepper noise. These noises will seriously interfere with the true information of the image, blurring or distorting the key state characteristics of the fruit surface, such as small rot spots, mold mycelium, color changes, and skin wrinkles. If a traditional overall smoothing filter method is used for noise reduction, although it can suppress noise to some extent, it often causes the edges and texture details of the image to be blurred together, resulting in the loss of key feature information. This confusion between noise and effective features directly reduces the accuracy of the subsequent feature extraction and state recognition model, ultimately affecting the decision-making accuracy and reliability of the entire preservation parameter optimization control system.

[0023] To solve the above technical problems, the present scheme proposes a two-stage noise reduction mechanism combining guided filtering and adaptive fusion, aiming to achieve accurate noise suppression and effective feature preservation.

[0024] The first stage of the mechanism is to perform a preliminary smoothing process with edge preservation characteristics on the original surface state image . In this way, the guided filter algorithm can smooth large homogeneous areas (such as the smooth skin of fruits) to filter out diffuse noise while using the gradient information of the image itself to protect high-frequency signal regions (such as the edges of fruit spots), avoiding the global blurring caused by traditional filtering. The execution process is as follows: the original image is used as the guide image , and for the local neighborhood centered on pixel in the image, the filter coefficients and are calculated by minimizing linear constraints. This process can be represented by the formula: Based on the calculated coefficients, a linear transformation is performed on all pixels in the neighborhood and the average is calculated to generate the output pixel , which ultimately forms the preliminary denoising image Q.

[0025] In this way, most of the random noise and impulse noise in the image is quickly and efficiently removed, providing a relatively "clean" reference image for subsequent refinement processing. Correspondingly, the significant noise points in the original image are effectively smoothed, and the overall visual quality of the image is preliminarily improved, but there is also a risk of slightly smoothing some very fine fruit surface texture that is misjudged as noise.

[0026] To make up for the possible over-smoothing problem in the first stage, the adaptive weighted fusion processing of the second stage is entered immediately. By quantifying the difference between the images before and after the preliminary filtering, it can be intelligently judged which areas are real noise and which areas are original details that should be preserved, so as to realize accurate correction of the noise reduction effect. The execution process is as follows: first, by calculating the pixel-level absolute difference between the original image P and the preliminary denoising image Q, a noise residual image N is generated. Expressed in formula as: The areas with higher values in the residual image N correspond to the filtered noise, while the areas with lower values correspond to the image structure that has not changed significantly. Subsequently, an adaptive weight map W is generated based on the residual image N, and the weight value of the weight map W in the strong noise area tends to 1, and the weight value in the detail feature area tends to 0. Finally, the original image P and the preliminary denoising image Q are weighted and fused using the weight map W to generate the final denoised fruit surface state image F.

[0027] In this way, the preliminary denoising result is refined and calibrated, and the effective details that may be over-smoothed in the first stage are selectively restored from the original image. The execution effect is that in the area where the noise is successfully filtered, the smoothed result after filtering is retained, while in the area containing key features, the clear details of the original image are maximally restored, and finally a fruit surface state image that is both clean and retains rich features is obtained.

[0028] Through the above technical means, the technical effect ultimately achieved by the above-mentioned noise reduction mechanism is that, while effectively removing the mixed noise in the fruit surface state image, the integrity and clarity of the edge, texture, and spot details and other detail features essential for fruit state judgment can be maintained with high fidelity. Compared with traditional noise reduction methods, the mechanism avoids simple compromises between noise reduction and detail preservation, achieving a dynamic balance between the two. It provides high-quality, high signal-to-noise ratio input data for subsequent fruit surface state multi-scale feature extractors. By outputting a clean and detailed fruit surface state image after noise reduction, the accuracy and robustness of subsequent feature extraction are greatly improved, ensuring that the fruit state recognizer can make accurate judgments based on true and reliable visual features, thereby providing precise data support for the entire fresh-keeping parameter adaptive optimization control system, and ensuring the scientificity and effectiveness of the fresh-keeping parameter adjustment such as oxygen concentration.

[0029] Specifically, the multi-scale feature extraction unit 132 is configured to input the fruit surface state image after noise reduction into a fruit surface state multi-scale feature extractor based on a dilated pyramid network to obtain the fruit state shallow feature map and the fruit state semantic feature map. It should be understood that, due to the complexity and diversity of the fruit surface state, single-scale feature extraction may not be able to fully reflect the surface state of the fruit. Therefore, the fruit surface state multi-scale feature extractor based on the dilated pyramid network is adopted in the present application to process the fruit surface state image after noise reduction. Specifically, the dilated pyramid network can extract the surface state feature information of the fruit from different scales by constructing a multi-scale feature pyramid, for example, by using a small-scale dilated convolution kernel to extract the detail features such as small defects and spots on the surface of the fruit, and by using a large-scale dilated convolution kernel to capture the overall maturity of the fruit and other extensive contextual semantic information, thereby forming multi-level fruit state representations, i.e., the fruit state shallow feature map and the fruit state semantic feature map.

[0030] In the above-mentioned fresh-keeping parameter adaptive optimization control system 100 based on fruit state detection, the fruit state shallow feature saliency module 140 is configured to perform shallow feature saliency on the fruit state shallow feature map in the parameter optimization control center to obtain a fruit state shallow salient feature map.

[0031] In a specific example of the present application, in the deep learning-based fruit state detection, although the shallow feature map of the neural network can effectively capture the basic visual elements such as edges, textures and colors of the image, it itself lacks discrimination. For an image of a fruit to be detected, the response intensity of the natural texture on the surface, the skin depression, and the early rotting spot, and the mold mycelium in the shallow feature map may have no essential difference, and even the response of the normal texture may be more complex and intense. This "one-size-fits-all" expression of all basic features results in the key diagnostic information being submerged in a large amount of background texture information. If the shallow feature map without screening and strengthening is directly used for subsequent analysis, the model is easy to misjudge the normal skin structure as a defect, or ignore the early lesion features that are not yet obvious, thereby causing deviation in the judgment of the true state of the fruit and reducing the sensitivity and specificity of the entire detection system.

[0032] To solve the above problems, the core of the present application is to analyze the structural information in the feature map, actively identify and amplify the features that constitute meaningful areas (fruit lesions), and guide the attention of the model.

[0033] The first step of this mechanism is to perform gradient field analysis and correlation distribution statistics on the shallow feature map. The reason for performing this step is that a meaningful area, such as a rotting patch, is often composed of a series of strong gradient features that are spatially adjacent and have inherent correlation in direction, while irregular background textures do not have this structural property. Therefore, by quantifying and statistically analyzing the gradient information, the two can be effectively distinguished. The execution process is as follows: first, the input fruit state shallow feature map is processed to calculate the gradient in the horizontal and vertical directions, and then the gradient amplitude and direction of each pixel point are synthesized. The formula for this step is: Then, strong gradient points with gradient amplitude higher than a certain threshold are selected, and the direction distribution of these points is counted to form a direction correlation distribution histogram of high-amplitude gradients The formula for this step is: The technical purpose of this step is not to directly modify the feature map, but to extract statistical descriptors that can represent structured information from the original shallow feature. In this way, the visual patterns in the shallow feature map are converted into a set of quantitative data (gradient amplitude and direction distribution), and in particular, a closed lesion will form a unique and identifiable distribution pattern in the direction histogram, providing a key basis for subsequent saliency judgment.

[0034] ​​The second step of the mechanism is to generate a saliency mask and apply it to the shallow feature map. The reason for this step is that statistical analysis alone is not enough to change the feature map itself. The "insights" gained from the analysis must be translated into an operational weight that can be applied to the original feature to truly highlight the key information. The process is as follows: First, identify the dominant directional distribution pattern that matches the fruit lesion shape (such as approximately circular or elliptical) from the directional association distribution histogram . Then, construct a saliency mask , which assigns higher values to pixel positions with high gradient intensity and gradient direction that matches the dominant pattern. After generating the preliminary mask, smoothing and normalization are usually performed to form a spatially continuous and appropriately valued weight map. Finally, apply this mask to the original shallow feature map in an additive enhancement manner to obtain the final fruit status shallow salient feature map. This step can be represented by the formula: The technical purpose of this step is to use the generated saliency mask as a magnifying glass to selectively enhance the original shallow feature map at the pixel level. The effect is that the activation values of features related to fruit lesions (whose gradient pattern was identified in the first step) are significantly improved, while features related to background texture remain essentially unchanged. This is like turning up the volume of the target sound in a noisy environment, making the target feature "stand out" and "easy to perceive" in the entire feature map.

[0035] In summary, the original shallow feature map, which contains a lot of irrelevant information, is transformed into a shallow salient feature map with key information significantly enhanced. It achieves a transition from "seeing everything" to "focusing on the key", greatly improving the signal-to-noise ratio of the shallow feature, where "signal" refers to features related to fruit status abnormalities, and "noise" refers to background texture with no diagnostic value. This provides a more pure and clear low-level visual input for subsequent feature fusion and advanced semantic analysis networks. By pre-highlighting the most likely lesion area, the attention of the entire deep learning model can be effectively guided, preventing it from being distracted by irrelevant details in the subsequent decision-making process. This not only improves the model's sensitivity to early and small lesions, but also enhances the accuracy of identifying different types of defects, thereby laying a solid data foundation for accurate fruit status recognition and reliable preservation parameter adaptive control.

[0036] In the above-mentioned fresh-keeping parameter adaptive optimization control system 100 based on fruit state detection, the fruit state recognition module 150 is configured to generate a recognition result of the fruit state based on the multi-scale fusion features between the fruit state shallow saliency feature map and the fruit state semantic feature map in the parameter optimization control center. Figure 4 FIG. 2 shows a block diagram of a fruit state recognition module according to an embodiment of the present application. Figure 4 As shown in FIG. 2, the fruit state recognition module 150 comprises a multi-scale feature joint perception unit 151 configured to input the fruit state shallow saliency feature map and the fruit state semantic feature map into a multi-scale feature joint perception network to obtain a fruit state multi-scale joint perception saliency feature map; and a recognition result generation unit 152 configured to input the fruit state multi-scale joint perception saliency feature map into a fruit state recognizer based on a classifier to obtain a recognition result, which is used to represent a fruit state label.

[0037] Specifically, the multi-scale feature joint perception unit 151 is configured to input the fruit state shallow saliency feature map and the fruit state semantic feature map into a multi-scale feature joint perception network to obtain a fruit state multi-scale joint perception saliency feature map.

[0038] In the deep learning analysis of the fruit state, different levels of neural networks produce feature maps with different properties. The shallow saliency feature map output by the shallow network retains rich spatial details, such as the accurate outline and texture of the rotting spots, but it lacks high-level semantic concepts and cannot distinguish between a spot being a lesion or a normal color spot of the fruit itself. In contrast, the semantic feature map output by the deep network contains highly abstract judgment information and can understand where the lesion may exist from a global perspective, but it has lost the accurate spatial positioning information and detailed texture in the repeated pooling and downsampling process. If these two feature maps with huge differences in scale and semantic level are simply spliced or added, it will cause serious semantic alignment conflicts: the high-level semantic information cannot accurately act on the low-level detailed features corresponding to it, resulting in the fused feature map being confused due to the introduction of irrelevant details, or the weight of the detailed features being improper, which weakens the basis for semantic judgment, and ultimately damages the recognition accuracy of the model for the fruit state, especially for early subtle lesions.

[0039] To solve the semantic gap problem in the multi-scale feature fusion, the present scheme constructs a multi-scale feature joint perception network, which uses a top-down guidance strategy to use high-level semantic information to accurately modulate and fuse shallow detailed features.

[0040] The first step of the mechanism is to generate a semantic-feature-based spatial attention weight map. The reason for this step is that the spatially directed information must be distilled from the semantic feature map containing high-level abstract concepts, i.e., to determine which regions are important in semantics.

[0041] The execution process is as follows: for the input fruit status semantic feature map, global average pooling and global maximum pooling are performed in parallel along its channel dimension to capture its global spatial context information from two dimensions and form two feature descriptors; after aggregating the two descriptors, a shared convolutional neural network layer is used to learn the complex nonlinear relationship between channels and map the result through an activation function (such as Sigmoid) to an initial spatial attention weight map; due to the low resolution of the semantic feature map, it is finally restored to the same spatial size as the shallow salient feature map through upsampling operation. The purpose of this step is to create a "semantic importance map" aligned with the high-resolution detail map, and each pixel value of the map quantifies the semantic saliency of the corresponding spatial position. The effect of this step is to generate a high-resolution weight mask, which presents high values in the area corresponding to the suspicious lesions of the fruit and low values in the background or normal area.

[0042] Subsequently, the mechanism enters the step of spatial attention weighting of shallow salient feature map under semantic guidance. The reason for this step is that the "semantic importance map" generated above must be used as a precise filter on the shallow feature map full of details but lacking focus. The execution process is as follows: the spatial attention weight map generated in the previous step is multiplied element by element with the input fruit status shallow salient feature map. This operation takes advantage of the value range characteristics of the attention weight map to amplify the shallow feature response corresponding to the area with high weight value (i.e., the area important in semantics), while suppressing the feature response corresponding to the area with low weight value (irrelevant background). The purpose of this step is to achieve selective enhancement of shallow detail features, ensuring that only those details confirmed by high-level semantics will be highlighted. The effect of this step is to output a semantic-guided shallow feature map, in which the edge and texture details related to the fruit lesion are effectively enhanced, while other irrelevant background texture details are significantly weakened.

[0043] Finally, the mechanism fuses the shallow features guided by semantics and the original semantic features. The reason for performing this link is that a complete feature expression needs both accurate details after screening and original, unweakened global context information as the basis for the final judgment. The execution process is as follows: first, the original fruit status semantic feature map is adjusted to the same spatial resolution as the shallow feature map after attention weighting through upsampling; then, the two are spliced in the channel dimension to form a deeper combined feature map containing two information sources; finally, the combined feature map is processed through one or several 1x1 convolution layers, aiming to learn the best fusion method of the two features through cross-channel information interaction and integrate and refine the feature dimension. The purpose of performing this link is to build a unified and efficient multi-scale feature expression. The effect is that a fruit status multi-scale joint perception saliency feature map is finally generated, which has both accurate detail description guided by high-level semantics and retains global abstract judgment information, realizing the cooperation and complementarity of information at different scales and semantic levels.

[0044] In this way, through the above series of technical means, it can effectively solve the fusion conflict between shallow detail features and deep semantic features in a deep neural network. Instead of simply forcing the two to be spliced, it establishes an attention guiding relationship from high-level semantics to low-level details, generating a high-quality fusion feature map that has both accurate spatial positioning and can reflect high-level judgments. In this way, a higher information density and stronger discriminative feature input is provided for the subsequent fruit status classifier. By ensuring that the model can pay attention to both "what defect" and "where the defect is and what the specific form is", the detection accuracy and robustness of the fruit surface state, especially early, fuzzy and irregular form lesions, are greatly improved. This provides a solid and reliable data foundation for the accurate decision of the entire preservation parameter self-adaptive optimization control system, ensuring that the system can respond quickly and effectively according to the most real fruit status changes.

[0045] Specifically, the recognition result generation unit 152 is configured to input the fruit status multi-scale joint perception saliency feature map into a fruit status recognizer based on a classifier to obtain a recognition result, which is used to represent a fruit status label. In the technical solution of the present application, the fruit status recognizer based on a classifier effectively learns the mapping relationship between fruit status features and predetermined labels through learning a large amount of training data. Based on this, when the fruit status multi-scale joint perception saliency feature map passes through the classifier, the classifier can use the network parameters learned in the training process to perform feature analysis and classification mapping on the fruit status multi-scale joint perception saliency feature map to obtain the corresponding fruit status label, such as "ripe", "unripe", "overripe", "skin damage", etc.

[0046] In the above-mentioned fresh-keeping parameter adaptive optimization control system 100 based on fruit state detection, the oxygen concentration adaptive adjustment module 160 is configured to generate an oxygen concentration adaptive adjustment instruction at the parameter optimization control center based on the identification result and the fruit state label-recommended oxygen concentration correspondence table. Specifically, the oxygen concentration adaptive adjustment module 160 is configured to: based on the identification result, match a recommended oxygen concentration value from the fruit state label-recommended oxygen concentration correspondence table; and based on a comparison between the recommended oxygen concentration value and a real-time oxygen concentration value collected by an oxygen sensor, generate an oxygen concentration adaptive adjustment instruction.

[0047] That is, after obtaining the fruit state label, the fruit state label-recommended oxygen concentration correspondence table is further queried and matched to extract a recommended oxygen concentration value matched with the fruit state label, and the oxygen concentration in the fruit storage environment is adjusted accordingly. That is, by comparing the recommended oxygen concentration value with the actual oxygen concentration value, an oxygen concentration adaptive adjustment instruction is generated. For example, when the actual oxygen concentration value is higher than the recommended value, a program for reducing the oxygen concentration is automatically started, and the oxygen concentration is diluted by injecting nitrogen or other inert gases until the recommended oxygen concentration value matched with the fruit state label is reached. Conversely, if the actual oxygen concentration value is lower than the recommended value, a program for increasing the oxygen concentration is started, and the oxygen concentration is increased by releasing pure oxygen to ensure that the fruit is stored in the best storage environment. In this way, dynamic and accurate regulation of the fruit fresh-keeping conditions is achieved, and the fruit is stored in the best storage conditions, thereby maximizing the shelf life and reducing economic losses and resource waste caused by improper fresh-keeping.

[0048] In summary, the fresh-keeping parameter adaptive optimization control system based on fruit state detection according to the embodiments of the present application is illustrated, which uses a camera to monitor the state of the fruit object and uses artificial intelligence technology based on deep learning to analyze the surface state image of the fruit object to extract multi-scale feature representations of the fruit surface state, and then uses the multi-scale feature representations to intelligently identify the fruit surface state class label. Meanwhile, a fruit state label-recommended oxygen concentration correspondence table obtained through experiments is used to determine the appropriate oxygen concentration value under the current state, thereby realizing adaptive optimization control of the oxygen concentration. In this way, the fresh-keeping effect can be effectively improved, the shelf life of the fruit can be extended, and losses caused by improper fresh-keeping can be reduced.

[0049] The basic principle of the present application is described above in combination with specific embodiments, however, it is pointed out that the advantages, benefits, effects and the like mentioned in the present application are only examples and are not limiting, and these advantages, benefits, effects and the like cannot be considered as necessary for each embodiment of the present application. In addition, the specific details of the above embodiments are only for the purpose of example and for the purpose of understanding, and the above details do not limit the present application to be necessarily implemented with the above specific details.

Claims

1. A fresh-keeping parameter adaptive optimization control system based on fruit status detection, characterized in that: include: A specific fruit object state monitoring module is used to collect surface state images of specific fruit objects through a camera; An image transmission module is used to input the surface state image of the specific fruit object into the parameter optimization control center, wherein the parameter optimization control center stores a fruit state label-recommended oxygen concentration correspondence table; A fruit state multi-scale feature extraction module is used to perform multi-scale feature extraction on the surface state image of the specific fruit object in the parameter optimization control center to obtain a fruit state shallow feature map and a fruit state semantic feature map; A fruit state shallow feature salient module is used to perform shallow feature salient on the fruit state shallow feature map in the parameter optimization control center to obtain a fruit state shallow salient feature map; A fruit state recognition module is used to generate a fruit state recognition result based on the multi-scale fusion features between the fruit state shallow salient feature map and the fruit state semantic feature map in the parameter optimization control center; The oxygen concentration adaptive adjustment module is used to generate an oxygen concentration adaptive adjustment instruction in the parameter optimization control center based on the recognition result and the fruit status label-recommended oxygen concentration correspondence table.

2. The adaptive optimization control system for preservation parameters based on fruit status detection according to claim 1 is characterized in that: The fruit state multi-scale feature extraction module includes: An image noise reduction unit, configured to perform noise reduction processing on the surface state image of the specific fruit object to obtain a noise-reduced surface state image of the fruit; The multi-scale feature extraction unit is used to input the denoised fruit surface state image into a fruit surface state multi-scale feature extractor based on a dilated pyramid network to obtain the fruit state shallow feature map and the fruit state semantic feature map.

3. The adaptive optimization control system for preservation parameters based on fruit status detection according to claim 2 is characterized in that: The image noise reduction unit is further configured to: Performing preliminary noise reduction processing based on guided filtering on the surface state image of the specific fruit object to obtain a preliminary noise reduction image; Adaptive weighted fusion based on noise assessment is performed on the preliminary denoised image and the surface state image of the specific fruit object to obtain the denoised fruit surface state image.

4. The adaptive optimization control system for preservation parameters based on fruit status detection according to claim 3 is characterized in that: The image noise reduction unit is further configured to: calculating a noise residual map between the preliminary denoised image and the surface state image of the specific fruit object; Normalizing the noise residual map to obtain an adaptive weight map; The preliminary denoised image and the surface state image of the specific fruit object are weightedly fused based on the adaptive weight map to obtain the denoised fruit surface state image.

5. The adaptive optimization control system for preservation parameters based on fruit status detection according to claim 4 is characterized in that: The image denoising unit is further configured to perform weighted fusion on the preliminary denoised image and the surface state image of the specific fruit object using the following formula to obtain a denoised fruit surface state image. The formula is: in, is the weight value of each position in the adaptive weight map, is the pixel value of each position in the surface state image of a specific fruit object, is the pixel value at each position in the preliminary denoised image.

6. The adaptive optimization control system for preservation parameters based on fruit status detection according to claim 2, characterized in that: The fruit state shallow feature salient module is used to: Perform gradient field analysis and correlation distribution statistics on the shallow feature map of the fruit state to obtain the gradient amplitude map and direction correlation distribution histogram; Based on the gradient amplitude map and the direction correlation distribution histogram, the shallow feature map of the fruit state is enhanced to obtain the shallow salient feature map of the fruit state.

7. The adaptive optimization control system for preservation parameters based on fruit status detection according to claim 1, characterized in that: The oxygen concentration adaptive adjustment module is used to: Based on the recognition result, matching the recommended oxygen concentration value from the fruit status label-recommended oxygen concentration correspondence table; Based on the comparison between the recommended oxygen concentration value and the real-time oxygen concentration value collected by the oxygen sensor, an oxygen concentration adaptive adjustment instruction is generated.

Images