A supply chain management method and system based on artificial intelligence

Abstract

The present invention relates to the field of image processing, and more specifically to an artificial intelligence-based supply chain management method and system. The method comprises: obtaining grayscale images of each side of goods when they are put into storage; selecting the grayscale image of any side as a target image, marking all initially identified areas in the target image, obtaining the intersection area of ​​all initially identified areas, calculating the prominence index of the intersection area, and forming the identification area set of the target image with the intersection area having a prominence index greater than a preset prominence threshold; inputting the identification area set of each target image of the goods into a neural network for training, and obtaining a multi-classification network after the training is completed; inputting the grayscale image of the goods collected in real time into the multi-classification network to obtain an output set containing probability values ​​of each category of the goods; and storing the goods in designated storage locations or shelves according to their categories based on the output set. The present invention can quickly identify the categories of goods and accurately store the goods.

Description

Technical Field

[0001] The present invention relates to the field of image processing. More specifically, the present invention relates to a supply chain management method and system based on artificial intelligence. Background Art

[0002] In supply chain management, warehousing is a crucial process. As the core hub of the supply chain, warehousing connects multiple links, including production, procurement, and sales. As a key activity in warehousing operations, the efficiency and accuracy of warehousing directly impact the operational efficiency and cost control of the entire supply chain. Efficient warehousing can shorten order processing cycles, reduce logistics costs, and improve customer satisfaction.

[0003] Prior art, such as the patent application document with publication number CN118840040A, discloses a three-dimensional warehouse system based on machine vision. The three-dimensional warehouse includes the following: First, an image acquisition unit captures an image of the goods to be stored. The pre-processing unit extracts the minimum circumscribed rectangle of the goods in the image and rotates the image of the rectangular area to be upright. The goods recognition unit uses a scale-invariant feature conversion algorithm to match the rotated image with template images of multiple known goods, and further determines the most matching goods template by calculating the pixel similarity, thereby identifying the category of the goods. Finally, the warehousing unit controls the warehousing equipment to store the goods in the corresponding allocable storage location in the three-dimensional warehouse based on the identified category of the goods.

[0004] However, the lighting in an actual warehouse environment is prone to change, and if a collision occurs during warehousing, the appearance of the existing goods will also change, thereby affecting the image quality collected by the image acquisition unit, resulting in the inability to perform efficient warehouse management later. Summary of the Invention

[0005] In order to solve the above-mentioned technical problem of the deterioration of the quality of cargo image acquisition and the impact on efficient warehouse management, the present invention provides solutions in the following aspects.

[0006] In a first aspect, a supply chain management method based on artificial intelligence includes:

[0007] Obtain grayscale images of each side of the goods when they enter the warehouse;

[0008] Selecting a grayscale image of any surface as a target image, marking all initially recognized areas in the target image, obtaining the intersection area of ​​all initially recognized areas, calculating the prominence index of the intersection area, and forming a recognition area set of the target image from the intersection areas having a prominence index greater than a preset prominence threshold; the recognition area set includes multiple cargo categories;

[0009] The set of recognition regions of each target image of the goods is fed into a neural network for training, and a multi-classification network is obtained after the training is completed; the grayscale image of the goods collected in real time is fed into the multi-classification network to obtain an output set containing probability values ​​of each category of the goods;

[0010] Based on the output collection, the goods are stored in the designated storage locations or shelves according to their categories.

[0011] The present invention accurately identifies the key characteristic regions of goods by acquiring grayscale images of each surface of the goods, marking the initial recognition areas in the target image, and then calculating the prominence index of the intersection area. These characteristic regions contain significant information about the goods' appearance, such as shape and texture. When these recognition region sets are fed into a neural network for training, the neural network learns the characteristic patterns of different goods categories and rapidly identifies the goods category. Furthermore, based on the probability values ​​of each goods category in the output set, the goods can be accurately stored in designated storage locations or shelves, making the storage of goods in the warehouse more orderly and allowing different categories of goods to be neatly placed together.

[0012] Preferably, the step of marking all initially identified areas in the target map includes:

[0013] Subtract the target image from all other grayscale images to obtain multiple difference images;

[0014] A center point is randomly selected on the difference map, and a circular area is constructed with the center point as the center and an initial radius of 0. Based on the preset expansion conditions, the initial radius is gradually increased to obtain circular areas of corresponding radius sizes. The recognition feature evaluation value of each circular area is calculated until the recognition feature evaluation value decreases for the first time, and the increase in radius is stopped. The circular area after the increase in radius is used as the initial recognition area of ​​the difference map, and then all the initial recognition areas in the target map are obtained.

[0015] By calculating the difference map and evaluating the local features, the important areas in the target image can be found adaptively, thereby improving the accuracy and robustness of image analysis.

[0016] Preferably, the protrusion index satisfies the relationship:

[0017] Where, is the prominence index of the intersection area, is the number of initial identification areas contained in the intersection area, is the mean value of all pixel values ​​in the intersection area, is the mean value of all pixel values ​​in the target image, is the hyperbolic tangent function.

[0018] By calculating and filtering the salience index, we can ensure that the target image's recognition area only contains areas with high salience. These areas are more likely to contain key features of the goods, such as shape, texture, and color, thus providing more accurate feature information for subsequent goods identification and classification.

[0019] Preferably, storing the goods by category in designated storage locations or shelves according to the output set includes:

[0020] A first classification threshold is set. If the maximum probability value of the goods category in the output set is greater than the first classification threshold, the input goods image is determined to belong to the goods category corresponding to the maximum probability value, and the corresponding goods are stored in the designated storage location or shelf according to the goods category.

[0021] By setting the first classification threshold, a clear judgment standard is provided for the classification of goods. Only when the maximum probability value of the goods category exceeds the first classification threshold, the goods are considered to belong to that category, thus reducing the possibility of misclassification.

[0022] Preferably, storing the goods by category in designated storage locations or shelves according to the output set further comprises:

[0023] Setting a classification threshold interval, and treating goods with the maximum probability value within the classification threshold interval as high-probability goods;

[0024] For each high-probability product, a judgment network is trained. The real-time product image is input into all trained judgment networks to obtain the output probability values ​​of multiple judgment networks. The judgment network with the maximum output probability value is found. The high-probability product corresponding to this judgment network is the product category to which the real-time collected product image belongs.

[0025] In actual classification, the probability value may approach the threshold, but there is still a certain degree of uncertainty. Setting a classification threshold interval can address these edge cases and avoid misclassification due to small probability differences. Training a judgment network for each high-probability item means that each item category has a corresponding judgment network. When a real-time item image is input into all trained judgment networks, the output probability values ​​of multiple judgment networks are obtained. By finding the judgment network with the maximum output probability value, the item category of the real-time item image is determined. This approach comprehensively considers the results of multiple judgment networks, reduces the error that may be caused by a single judgment network, and further improves classification accuracy.

[0026] Preferably, the training process of the judgment network includes:

[0027] A second classification threshold is set, and goods with a maximum probability value less than the second classification threshold are regarded as low-probability goods; for each high-probability goods, the area containing only low-probability goods is removed from its recognition area set to obtain a new recognition area set; in the new recognition area set, if any recognition area contains high-probability goods, the maximum probability value corresponding to the high-probability goods is added by 1 as the label value of the recognition area, otherwise 1 is used as the label value of the recognition area; all label values ​​are weighted to obtain the weight of each recognition area in the new recognition area set; each recognition area in the new recognition area set is encoded to obtain a corresponding encoding vector, all encoding vectors are accumulated to obtain a discriminant vector, and training is performed based on the discriminant vector to obtain a judgment network for high-probability goods.

[0028] For the new set of identification areas, different tag values ​​are assigned to each identification area based on whether the identification area contains high-probability goods, thereby highlighting the characteristics of high-probability goods and making the classification results more accurate; further, the characteristic information of high-probability goods is extracted through the encoding and accumulation process. Based on the judgment network of high-probability goods, the goods can be more accurately classified and stored in designated storage locations or shelves, making the layout of goods in the warehouse more reasonable, and different categories of goods can be neatly placed together for easy management and search.

[0029] Preferably, a cross entropy loss function is used in the judgment network training process, and the training of the judgment network is completed when a preset number of training times is reached or the loss function value is less than a preset loss threshold.

[0030] In a second aspect, a supply chain management system based on artificial intelligence includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned supply chain management method based on artificial intelligence is implemented.

[0031] The beneficial effects of the present invention are:

[0032] The present invention collects cargo images in real time and processes and analyzes them, classifying and storing cargo in a timely manner. At the same time, when faced with new cargo types or changing cargo characteristics, the present invention continuously trains and optimizes neural networks to quickly adapt and accurately identify them, thereby maintaining the dynamic and flexible nature of supply chain management. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an illustrative and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0034] Figure 1This is a method flow chart of steps S1 to S4 in an artificial intelligence-based supply chain management method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention.

[0036] The application scenario of the present invention is to perform image analysis on six-sided rectangular parallelepiped goods as an example.

[0037] The embodiment of the present invention discloses a supply chain management method based on artificial intelligence, referring to Figure 1 , including steps S1 to S4, specifically as follows:

[0038] S1: Obtain grayscale images of each surface of the goods when they enter the warehouse.

[0039] The appearance of goods can vary depending on factors such as placement angle and lighting conditions. Collecting a grayscale image of only one side may not provide a clear visual representation of the goods' characteristics. However, capturing grayscale images of all sides provides a more comprehensive picture of the goods' appearance, improving identification accuracy. Furthermore, goods vary in shape, size, and material. Some goods may have multiple similar sides, while others may each have unique characteristics.

[0040] Specifically, place the goods within the shooting range of the image acquisition device to ensure that the goods are stable and avoid movement or tilting during the shooting process. A conveyor belt or fixing device can be used to assist in positioning the goods.

[0041] Photograph each side of the cargo in a pre-set order. For example, for a common six-sided rectangular cargo, you need to photograph all six sides: the top, bottom, front, back, left, and right. When photographing each side, ensure that the lens is perpendicular to the surface of the cargo to obtain a clear, distortion-free image.

[0042] The original images collected above are converted into grayscale images.

[0043] S2: Select a grayscale image of any surface as a target image, mark all initially identified areas in the target image, obtain the intersection area of ​​all initially identified areas, calculate the prominence index of the intersection area, and form the intersection areas with prominence index greater than a preset prominence threshold into a recognition area set of the target image; the recognition area set includes multiple cargo categories.

[0044] In practical applications, the image acquisition process may be affected by various factors such as lighting, shadows, and reflections, resulting in degraded image quality. By generating a difference map, we can eliminate these interfering factors to a certain extent, as they may be weakened or eliminated in the difference map.

[0045] Furthermore, since goods may be placed at any angle when entering the warehouse, it is necessary to consider the recognition of images at different rotation angles. By generating difference maps at multiple rotation angles, the algorithm's robustness to rotation can be improved. Even if the angle of the goods changes, the goods can still be identified by comparing the difference maps.

[0046] Specifically, a grayscale image of any side is used as the target image for analysis. For the target image, pixel values ​​are subtracted from the images of all other corresponding sides of the goods at five rotation angles (such as 0 degrees, 90 degrees, 180 degrees, 270 degrees, and 360 degrees; although 360 degrees is equal to the original image, it is included in the rotation angle for completeness) to obtain five difference images.

[0047] After obtaining the difference map of the target image, there may be multiple distinctive feature areas in the difference map. These features may be trademarks, labels, special patterns or textures, etc. These features are crucial for distinguishing different goods.

[0048] Specifically, for a difference map, a center point is randomly selected, and a circular area is constructed with the center point as the center and an initial radius of 0. Then, the initial radius is gradually increased by one pixel unit distance to expand the circular area. Initially, the radius is small, and the pixel differences in the circular area may be relatively consistent. As the radius increases, the pixel differences in the area gradually contain more difference information.

[0049] After each expansion, the recognition feature evaluation value of the corresponding circular area is calculated. When the recognition feature evaluation value decreases for the first time, the expansion of the circular area radius is stopped, and the circular area after the expansion is stopped is used as the initial recognition area of ​​the difference map.

[0050] The recognition feature evaluation value of the circular area is calculated by analyzing the distribution of pixel values ​​within the circular area, that is, the relationship is satisfied:

[0051]

[0052] Where, Expand the radius of the circular area The corresponding recognition feature evaluation value of the circular area is Expand the radius of the circular area Pixel difference The probability of occurrence, Expand the radius of the circular area The mean of the absolute values ​​of all pixel differences at this time, Expand the radius of the circular area The corresponding value at this time is is the hyperbolic tangent function.

[0053] in, The entropy used to calculate the probability distribution of pixel differences within a circular area. The larger the entropy, the more uniform the distribution of pixel differences, the higher the amount of information in the area, and the higher the degree of distinction. Normalize the mean of the absolute values ​​of all pixel differences in the circular area so that the feature evaluation varies within a certain range; The radius of the circular area is normalized so that the feature evaluation is not only related to the distribution of pixel differences in the area, but also to the size of the area, avoiding feature evaluation deviation caused by the area being too large or too small.

[0054] As the radius increases, the pixel difference within the area gradually contains more difference information, and the recognition feature evaluation also changes accordingly. When the recognition feature evaluation decreases for the first time, it means that the pixel value difference information within the circular area has reached saturation. Continuing to expand the area will not bring more effective feature information, but may introduce noise or irrelevant information. Stopping the expansion at this time can ensure that the initial recognition area has high feature significance and discrimination, providing a reliable basis for subsequent cargo identification.

[0055] If the pixel differences within a circular area are uniformly distributed and consistent, the identification feature evaluation value may be low. Conversely, if the pixel differences within a circular area show large fluctuations, variations, or heterogeneity (for example, containing multiple colors, textures, or brightness variations), the identification feature evaluation value may be high. A high feature evaluation value may mean that the circular area contains more information or details.

[0056] The method of determining the initial recognition area of ​​the above difference map can be similarly used to obtain the initial recognition areas corresponding to other difference maps.

[0057] In summary, one target image corresponds to five difference images, each difference image corresponds to one preliminary recognition region, and thus all preliminary recognition regions in the target image are obtained.

[0058] Traverse all the marked initial identification areas and detect the overlap between them. For each pair of areas, calculate their overlapping parts. This can be achieved using geometric algorithms or functions in the image processing library.

[0059] For each pair of overlapping initial recognition areas, their overlapping parts are treated as a new intersection area. If multiple areas overlap, their overlapping parts can be merged into a larger intersection area. For example, if there are 8 initial recognition areas and there are two initial recognition areas that overlap, then there is 1 intersection area, and the number of initial recognition areas that intersect in this intersection area is 2; or if there are three initial recognition areas that overlap, then there is 1 intersection area, and the number of initial recognition areas that intersect in this intersection area is 3.

[0060] All intersection areas of the target map are counted, and the significance and discrimination of the intersection areas are further analyzed by calculating the prominence index of the intersection areas.

[0061] Specifically, the number of initially identified areas contained in the intersection area, the mean of all pixel values ​​in the intersection area, and the mean of all pixel values ​​in the target image are obtained. The prominence index of the intersection area is calculated based on the number of initially identified areas contained in the intersection area, the mean of all pixel values ​​in the intersection area, and the mean of all pixel values ​​in the target image, that is, the relationship is satisfied:

[0062]

[0063] Where, is the prominence index of the intersection area, is the number of initial identification areas contained in the intersection area, is the mean value of all pixel values ​​in the intersection area, is the mean value of all pixel values ​​in the target image, is the hyperbolic tangent function.

[0064] in, The relative difference between the average pixel value of the intersection area and the average pixel value of the entire target image is calculated. If the pixel values ​​of the intersection area and the entire target image are very close, then this value will be close to 0; if the difference is large, this value will be close to 1.

[0065] By combining the number of initially identified regions and the pixel value difference between the intersection region and the entire target image, this index provides a quantitative indicator to assess the significance of a region in the target image. A high value indicates that the region is more prominent or significant in the target image, perhaps because it contains multiple initially identified regions and its pixel values ​​are significantly different from the background.

[0066] Furthermore, the prominence indexes of all intersection areas of the target graph are calculated, and the top 10 intersection areas with the largest prominence indexes are retained (it should be noted that in actual applications, the number of intersection areas to be retained can be dynamically adjusted according to the actual situation of the goods and the recognition requirements). The corresponding intersection areas constitute the recognition area set of the target graph. This means that for each target graph, only those areas that are most likely to provide useful recognition information are focused on.

[0067] In summary, for each target graph of the goods, 10 intersection areas with higher prominence indexes are selected as the recognition area set, so 60 recognition areas can be obtained by adding them up.

[0068] Furthermore, in these recognition areas, there may be overlapping images or data of multiple categories of goods (i.e., intersections). The categories of goods that can be distinguished in these recognition areas are recorded. This means that if a recognition area contains image features of two or more types of goods at the same time, for example, if a recognition area appears in images of multiple goods, and these goods belong to different categories, then this recognition area contains multiple categories of goods.

[0069] In general, the above S2 operation utilizes the differences and significance of image features, and constructs an effective recognition area set by calculating and screening the intersection areas with larger prominence indexes, thereby helping to distinguish and identify different categories of goods.

[0070] S3: The recognition area set of each target image of the goods is put into the neural network for training. After the training is completed, a multi-classification network is obtained; the grayscale image of the goods collected in real time is put into the multi-classification network to obtain an output set containing the probability value of each goods category.

[0071] In real-world applications, cargo images may be affected by a variety of factors, such as lighting variations, image quality, and occlusion. Through training, neural networks learn the features within the recognition area, which are crucial for distinguishing between different categories of objects. During training, the neural network continuously adjusts its internal parameters (such as weights and biases) to produce the desired output for a given input. This learning process enables the neural network to gradually improve its recognition accuracy for the recognition area.

[0072] Specifically, the set of recognition regions of each target image of the goods is put into a neural network for training, and a multi-classification network is obtained after the training is completed. It should be noted that neural network training and multi-classification networks are existing technologies and will not be described in detail here.

[0073] According to the operation of S1 above, the grayscale image of the goods collected in real time is input into the multi-classification network to obtain an output set containing the probability values ​​of each goods category.

[0074] S4: Based on the output set, the goods are stored in designated storage locations or shelves according to their categories.

[0075] First, a first classification threshold is set (set to 0.8 in the embodiment of the present invention). The first classification threshold is used to distinguish whether the cargo image can be clearly classified into a certain cargo category.

[0076] Then, the maximum probability value of the cargo category is found in the output set obtained in the above S3. If the maximum probability value is greater than the first classification threshold, it is considered that the cargo image can be clearly classified into the cargo category corresponding to the maximum probability value.

[0077] Next, according to this cargo category, the corresponding cargo is stored in the designated storage location or shelf.

[0078] It's important to consider that in classification tasks, a threshold is typically set to determine the reliability of the prediction results. A common threshold is 0.5, but in this embodiment, a more stringent threshold of 0.8 is selected. In this case, if the maximum probability value falls within the classification threshold interval of [0.5, 0.8], this means that while the probability for that category is the highest, it still falls short of the set threshold, and therefore, it cannot be determined that the input product image belongs to that category.

[0079] In addition, for the judgment of not belonging to a certain category, this threshold is usually relatively low. For example, if the threshold is set to 0.2, then when the probability value of a certain category is lower than this threshold, it can be considered that the product image does not belong to this category. Therefore, the second classification threshold is set to 0.2.

[0080] Therefore, in order to improve the accuracy of judgment, the goods corresponding to the maximum probability value of the goods category between [0.5, 0.8] are regarded as high-probability goods, and the goods corresponding to the maximum probability value of the goods category less than 0.2 are regarded as low-probability goods.

[0081] For each high-probability cargo, the area containing only low-probability cargo categories is removed from its identification area set to obtain a new identification area set. In the new identification area set, if an identification area contains high-probability cargo, the maximum probability value corresponding to the high-probability cargo is added by 1 as the mark value of the identification area. If the identification area does not contain high-probability cargo, 1 is used as the mark value of the identification area. Then, all the mark values ​​are weighted to obtain the weight of each identification area in the new identification area set.

[0082] Each identification area in the new identification area set is encoded to obtain a corresponding encoding vector. All encoding vectors are accumulated to obtain a discriminant vector. This discriminant vector is then used for training to form a network for identifying high-probability items. The network training process uses a cross-entropy loss function. Training is completed when a preset number of training cycles is reached or the loss function value falls below a preset loss threshold (in this embodiment, the loss threshold is set to 0.5).

[0083] Specifically, the above discriminant vector satisfies the relationship:

[0084]

[0085] Where, is the discriminant vector of high probability goods, New identification area for high probability goods The weight of the recognition area, New identification area for high probability goods The encoding vector of the identified region, The total number of new identification areas in the concentrated identification area for high-probability goods.

[0086] The above S4 screens out goods with certain and uncertain classifications through a preliminary multi-classification network, and further trains a special judgment network for the uncertain goods to improve the accuracy of classification.

[0087] The present invention can accurately determine the key feature areas of the goods by obtaining grayscale images of each side of the goods, marking the initial recognition areas in the target image, and then calculating the prominence index of the intersection area. These feature areas contain significant information about the appearance of the goods, such as shape, texture, etc. After these recognition area sets are put into the neural network for training, the neural network can learn the characteristic patterns of different categories of goods. For example, in a logistics warehouse, there are various kinds of goods, such as electronic product packaging boxes and packaging boxes. This process can quickly identify the regular shape of the packaging box and features such as metal accessories on the packaging box, so as to quickly determine the category of the goods when the goods are put into the warehouse, avoiding the tedious process of manually checking the goods one by one, and greatly improving the recognition efficiency.

[0088] An embodiment of the present invention further discloses an artificial intelligence-based supply chain management system, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, an artificial intelligence-based supply chain management method according to the present invention is implemented.

[0089] The system also includes other components well known to those skilled in the art, such as a communication bus and a communication interface. The configuration and functions of these components are known in the art and therefore will not be described in detail here.

[0090] In the present invention, the aforementioned memory can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic storage medium or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store the required information and can be accessed by an application, module, or both. Any such computer storage medium can be part of, accessible to, or connected to the device.

[0091] In the description of this specification, "multiple" and "several" mean at least two, such as two, three or more, unless otherwise clearly defined.

[0092] While several embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Numerous modifications, variations, and alternatives will occur to those skilled in the art without departing from the concept and spirit of the present invention. It should be understood that various alternatives to the embodiments of the present invention described herein may be employed in practicing the present invention.

Claims

1. A supply chain management method based on artificial intelligence, characterized in that: include: Obtain grayscale images of each side of the goods when they enter the warehouse; A grayscale image of any surface is selected as a target image, and the target image is subtracted from all other grayscale images to obtain multiple difference images; a center point is randomly selected on the difference image, and a circular area is constructed with the center point as the center and an initial radius of 0. The initial radius is gradually increased based on a preset expansion condition to obtain circular areas of corresponding radius sizes, and the recognition feature evaluation value of each circular area is calculated. The increase in radius is stopped until the recognition feature evaluation value decreases for the first time, and the circular area after the increase in radius is stopped is used as the initial recognition area of ​​the difference image, thereby obtaining all the initial recognition areas in the target image; the intersection area of ​​all the initial recognition areas is obtained, and the prominence index of the intersection area is calculated. The intersection areas with prominence indexes greater than a preset prominence threshold are used to form a recognition area set of the target image; the recognition area set includes multiple cargo categories; The set of recognition regions of each target image of the goods is fed into a neural network for training, and a multi-classification network is obtained after the training is completed; the grayscale image of the goods collected in real time is fed into the multi-classification network to obtain an output set containing probability values ​​of each category of the goods; Based on the output collection, the goods are stored in the designated storage locations or shelves according to their categories.

2. The supply chain management method based on artificial intelligence according to claim 1, characterized in that: The protrusion index satisfies the relationship: Where, is the prominence index of the intersection area, is the number of initial identification areas contained in the intersection area, is the mean value of all pixel values ​​in the intersection area, is the mean value of all pixel values ​​in the target image, is the hyperbolic tangent function.

3. The supply chain management method based on artificial intelligence according to claim 2, characterized in that: According to the output set, storing the goods by category in designated storage locations or shelves includes: A first classification threshold is set. If the maximum probability value of the goods category in the output set is greater than the first classification threshold, the input goods image is determined to belong to the goods category corresponding to the maximum probability value, and the corresponding goods are stored in the designated storage location or shelf according to the goods category.

4. The supply chain management method based on artificial intelligence according to claim 3, characterized in that: The storing of goods by category in designated storage locations or shelves according to the output set further includes: Setting a classification threshold interval, and treating goods with the maximum probability value within the classification threshold interval as high-probability goods; For each high-probability product, a judgment network is trained. The real-time product image is input into all trained judgment networks to obtain the output probability values ​​of multiple judgment networks. The judgment network with the maximum output probability value is found. The high-probability product corresponding to this judgment network is the product category to which the real-time collected product image belongs.

5. The supply chain management method based on artificial intelligence according to claim 4, characterized in that: The training process of the judgment network includes: A second classification threshold is set, and goods with a maximum probability value less than the second classification threshold are regarded as low-probability goods; for each high-probability goods, the area containing only low-probability goods is removed from its recognition area set to obtain a new recognition area set; in the new recognition area set, if any recognition area contains high-probability goods, the maximum probability value corresponding to the high-probability goods is added by 1 as the label value of the recognition area, otherwise 1 is used as the label value of the recognition area; all label values ​​are weighted to obtain the weight of each recognition area in the new recognition area set; each recognition area in the new recognition area set is encoded to obtain a corresponding encoding vector, all encoding vectors are accumulated to obtain a discriminant vector, and training is performed based on the discriminant vector to obtain a judgment network for high-probability goods.

6. The supply chain management method based on artificial intelligence according to claim 5, characterized in that: The cross entropy loss function is used in the judgment network training process. When the preset number of training times is reached or the loss function value is less than the preset loss threshold, the judgment network training is completed.

7. A supply chain management system based on artificial intelligence, characterized in that: include: A processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the supply chain management method based on artificial intelligence according to any one of claims 1 to 6 is implemented.

Images