Commodity appearance image processing method for large-scale commodity library

Abstract

The invention discloses a commodity appearance image processing method for a large-scale commodity library, and relates to the technical field of image processing. When a new commodity needs to enter a commodity library, a new commodity image is acquired and preprocessed, and then color features, texture features and shape features are extracted and input into a classification model for comprehensive evaluation, so that the category deviation index of the new commodity and various commodities in the classification model is calculated; according to the method, the category corresponding to the highest index is selected as the estimated category, accurate classification can be carried out based on the image features of the new commodities, orderly management of the commodities in a commodity library is facilitated, and the problems that in the prior art, a traditional processing mode for massive commodity images is low in efficiency, and accuracy cannot be guaranteed when the commodities enter the commodity library are solved.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a method for processing product appearance images for a large-scale product library. Background Art

[0002] With the rapid development of e-commerce, the establishment and management of large-scale product libraries have become crucial. In product libraries, product appearance images are important elements for displaying product information and attracting consumers.

[0003] However, the existing product appearance image processing methods still have the following shortcomings in practical applications: Traditional processing methods are inefficient for massive amounts of product images, and their accuracy cannot be guaranteed when they enter the product library. On the other hand, it is inefficient in processing massive amounts of product images, making it difficult to meet the needs of rapid updates and retrieval, and it is not possible to dynamically select better nodes for processing.

[0004] Therefore, a product appearance image processing method for large-scale product libraries is proposed. Summary of the invention

[0005] In order to solve the problems raised in the background technology, the present application provides a method for processing product appearance images for a large-scale product library.

[0006] The product appearance image processing method for a large-scale product library provided in the present application adopts the following technical solution: The product appearance image processing method for a large-scale product library comprises: Establish a classification model: collect image samples of various categories of commodities stored in the commodity library, and preprocess the image samples of various categories of commodities; after the preprocessing is completed, extract the color features, texture features and shape features of the image samples of various categories of commodities stored in the commodity library, and input the extracted features into the model structure according to the pre-selected model architecture to establish the classification model corresponding to the current commodity library; Input product image: When a new product needs to enter the product library, obtain the new product image and extract color features, texture features and shape features after preprocessing, and input it into the classification model for comprehensive evaluation, so as to output the category bias index tri between the current new product and each category of products in the classification model; where i represents the number of each category of products, i=1, 2, ..., k, k is the total number of product categories; Output product category: According to the category bias index tri between the current new product and each category of products in the classification model, select the category corresponding to the highest category bias index tri from each group of category bias index tri as the estimated category of the current new product, and transport the new product to the storage location of the corresponding estimated category in the product library through automated equipment.

[0007] Optionally, obtain new product images, perform preprocessing, extract color features, texture features, and shape features, and input them into the classification model for comprehensive evaluation, specifically: S1: For the color features of the new product image and the samples of each type of product image in the classification model, count the number of pixels in each color interval of the new product image and the samples of each type of product image to obtain the color histogram of the new product image and the samples of each type of product image; Calculate the Bhattacharyya distance between the new product image and the color histogram of each type of product image sample as the color similarity index mai between each type of product image sample and the new product image; S2: Generate gray-level co-occurrence matrices for the texture features of the new product image and the various product image samples in the classification model; extract contrast, correlation, energy, and entropy features from the gray-level co-occurrence matrix, and thereby construct texture feature vectors for the new product image and the various product image samples in the classification model; Calculate the Mahalanobis distance between the texture feature vectors of the new product image and the samples of various product images as the texture similarity index mbi between the samples of various types of images and the new product image; S3: For the shape features of the new product image and the various product image samples in the classification model, extract the contours of the new product image and the various product image samples, and calculate the perimeter, area, circularity and eccentricity of the contours, thereby constructing shape descriptors of the new product image and the various product image samples in the classification model; The Euclidean distance is used to calculate the distance between the new product image and the shape descriptors of various product images, which is used as the shape similarity index mci between various image samples and the new product image.

[0008] Optionally, output the category bias index tri between the current new product and each category of products in the classification model, specifically: S4: preset the color reference similarity index, texture reference similarity index and shape reference similarity index, and mark them as yai, ybi and yci; extract the color similarity index mai, texture similarity index mbi and shape similarity index mci of the new product image and various image samples, and substitute them into the formula Perform weighted calculation to obtain the category bias index tri between the current new product and the various products in the classification model; , as well as They are the influencing weight factors of color similarity index mai, texture similarity index mbi and shape similarity index mci respectively.

[0009] Optionally, also include: Product processing optimization: The constructed classification model is deployed on each node for parallel processing. When a new product needs to enter the product library, the load parameters of each node are analyzed to obtain the load evaluation index fzp of each node, and the node with the lowest load evaluation index fzp is selected as the allocation node for the new product image; the load parameters include task queue length, average task processing time, CPU usage, and memory usage.

[0010] Optionally, analyze the load parameters of each node, specifically: Get the number of new product images waiting to be processed at each node at the current time point, so as to obtain the task queue length of each node; Obtain the processing time used by each node to process E groups of new product images before the current time point; thereby calculate the average task processing time of each node at the current time point, which is calculated by dividing the processing time used by E; Get the CPU usage and memory usage of each node at the current time point, set the weight coefficients corresponding to the CPU usage and memory usage, multiply the CPU usage and memory usage of each node at the current time point by the corresponding set weight coefficients, and then sum them up to get the computing status index of each node at the current time point.

[0011] Optionally, obtain the load evaluation index fzp of each node, specifically: Preset the length value range, duration value range and index value range of each group corresponding to the task queue length, average task processing time and operation status index respectively; set each group of length value range to correspond to a task load score, each group of duration value range to correspond to a node timeliness score, and each group of index value range to correspond to a node load score; Match the task queue length, average task processing time and operation status index of each node at the current time point with the corresponding length value range, time value range and index value range of each group, so as to determine the task load score, node timeliness score and node load score of each node at the current time point; Set the weight coefficients corresponding to the task load score, node timeliness score and node load score respectively, multiply the task load score, node timeliness score and node load score of each node at the current time point by the corresponding set weight coefficients, and then sum them up to obtain the load evaluation index fzp of each node at the current time point.

[0012] Optionally, also include: Classification model verification: After the classification model corresponding to the current product library is established, according to the set evaluation time interval, after the set evaluation time interval is reached, X product images are randomly extracted from each type of product image in the pre-built image database; and input into the classification model for verification, so as to obtain the classification effect evaluation index yu of the classification model at the current time point, and according to the classification effect evaluation index yu of the classification model at the current time point, selectively trigger the adjustment signaling and send it to the technical staff.

[0013] Optionally, according to the classification effect evaluation index yu of the classification model at the current time point, the adjustment signaling is selectively triggered and sent to the technical staff, specifically: Input the extracted X product images of various categories into the established classification model, and output the estimated categories corresponding to the X product images; The estimated categories corresponding to the X product images of each category are matched with the extracted category areas of the image database. If the match is successful, it is judged as a correct classification, otherwise it is judged as an abnormal classification; For the X product images extracted from each category, the percentage of correct classifications in each category is counted, that is, calculated by the number of correct classifications / X, and the classification accuracy rate fai of the classification model for each category of products is obtained; Preset the passing accuracy rate of various types of goods and mark them as fbi; According to the formula The classification accuracy fai of various types of commodities is weighted to obtain the classification effect evaluation index yu of the classification model at the current time point; is the influencing weight factor of the classification accuracy fai of various types of goods; The classification effect evaluation index yu of the classification model at the current time point is compared with the preset effect threshold index. If the classification effect evaluation index yu of the classification model at the current time point is less than the preset effect threshold index, the adjustment signal is triggered and sent to the technician.

[0014] In summary, the present application includes at least one of the following beneficial technical effects: When a new product needs to enter the product library, the present invention obtains the new product image and extracts the color features, texture features and shape features after preprocessing, and inputs them into the classification model for comprehensive evaluation, thereby calculating the category bias index of the new product and various types of products in the classification model, and selecting the category corresponding to the highest index as the estimated category. Based on the image features of the new product, a relatively accurate classification can be performed, which helps the product library to manage the products in an orderly manner, and solves the problem in the prior art that the traditional processing method for massive product images is inefficient and the accuracy cannot be guaranteed when entering the product library; The present invention deploys the classification model to multiple nodes for parallel processing, analyzes load parameters such as task queue length, average task processing time, CPU usage and memory usage, calculates the load evaluation index, and distributes new product images to the nodes with the lowest load, thereby avoiding task backlogs, improving processing efficiency, and realizing rapid processing and retrieval of product images. After establishing the classification model, the present invention sets the evaluation time interval, extracts images for verification from a dynamically updated image database, calculates the classification effect evaluation index, and compares it with a preset threshold. If it is lower than the threshold, an adjustment signal is triggered to notify the technician, so that the model can be optimized in time to ensure the stable operation of the classification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Further details, features and advantages of the present application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0016] Several embodiments of the present application will be described in more detail below with reference to the accompanying drawings so that those skilled in the art can implement the present application. The present application can be embodied in many different forms and purposes and should not be limited to the embodiments described herein. These embodiments are provided to make the present application comprehensive and complete, and to fully convey the scope of the present application to those skilled in the art. The embodiments do not limit the present application.

[0017] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those of ordinary skill in the art to which this application belongs. It will be further understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the relevant art and / or the context of this specification, and will not be interpreted in an idealized or overly formal sense unless explicitly defined as such herein.

[0018] See also Figure 1 As shown, the product appearance image processing method for a large-scale product library includes: Establish a classification model: collect image samples of various categories of goods stored in the product library, and preprocess the image samples of various categories of goods; the preprocessing includes denoising, grayscale and normalization operations; after the preprocessing is completed, perform feature extraction on the preprocessed image samples of various categories of goods, extract the color features, texture features and shape features of the image samples of various categories of goods stored in the product library, and input the extracted features into the model structure according to the pre-selected model architecture to establish the classification model corresponding to the current product library; the pre-selected model architecture is a convolutional neural network; It should be noted that the process of establishing a classification model, for example: Collect clothing and electronic product image samples from the product library, for example, collect clothing images of different styles of tops, pants, skirts, etc., and electronic product images such as mobile phones, computers, headphones, etc.; Gaussian filtering is used to denoise the image, convert the color image into a grayscale image, and normalize the pixel value of the image to the interval [0, 1]; Calculate the color histogram of the image. For clothing images, count the distribution ratios of different colors in the image, such as the proportion of red, blue, green, etc., and use this as the color feature; The texture features are extracted using the gray-level co-occurrence matrix. For electronic product images, the texture features such as contrast, correlation, energy and entropy are calculated by GLCM to reflect the spatial distribution relationship of the gray levels in the image. The contour of the product is obtained through edge detection, and then the shape descriptors such as the perimeter, area, and circularity of the contour are calculated. For example, for a skirt image, the shape of the skirt can be described by shape features. Determine the size of the input layer according to the size of the preprocessed image, add multiple convolutional layers, add a pooling layer after the convolutional layer, flatten the pooled feature map, and connect it to the fully connected layer; the number of neurons in the output layer is equal to the number of product categories, build a convolutional neural network model, and use the built convolutional neural network model as the classification model corresponding to the current product library; Input product image: When a new product needs to enter the product library, obtain the new product image and extract color features, texture features and shape features after preprocessing, and input it into the classification model for comprehensive evaluation, so as to output the category bias index tri between the current new product and each category of products in the classification model; where i represents the number of each category of products, i=1, 2, ..., k, k is the total number of product categories; Specifically: S1: For the color features of the new product image and the samples of each type of product image in the classification model, count the number of pixels in each color interval of the new product image and the samples of each type of product image to obtain the color histogram of the new product image and the samples of each type of product image; Calculate the Bhattacharyya distance between the color histograms of the new product image and the samples of each type of product image as the color similarity index mai between the samples of each type of product image and the new product image; the smaller the Bhattacharyya distance, the more similar the two color histograms are, that is, the more similar the color features of the new product image are to the images of this type of product; S2: For the texture features of the new product image and the various product image samples in the classification model, the gray level co-occurrence matrix is ​​generated respectively; the gray level co-occurrence matrix describes the probability of different gray level pairs in the image appearing simultaneously at a certain spatial distance and direction; the contrast, correlation, energy and entropy features are extracted from the gray level co-occurrence matrix, and the texture feature vectors of the new product image and the various product image samples in the classification model are constructed; the contrast reflects the intensity of the local gray level change in the image; the correlation measures the linear correlation between the gray level pairs; the energy represents the uniformity of the gray level distribution of the image; the entropy measures the complexity of the texture in the image; Calculate the Mahalanobis distance between the texture feature vectors of the new product image and the samples of each type of product image as the texture similarity index mbi of each type of image sample and the new product image; the smaller the Mahalanobis distance, the more similar the texture features of the new product image are to the images of this type of product; S3: For the shape features of the new product image and the various product image samples in the classification model, extract the contours of the new product image and the various product image samples, and calculate the perimeter, area, circularity and eccentricity of the contours, thereby constructing shape descriptors of the new product image and the various product image samples in the classification model; The Euclidean distance is used to calculate the distance between the new product image and the shape descriptors of various product images, which is used as the shape similarity index mci of various image samples and the new product image. The smaller the Euclidean distance, the more similar the shape features of the new product image are to the image of the product of that category. S4: preset color reference similarity index, texture reference similarity index and shape reference similarity index, and mark them as yai, ybi and yci; set by technical personnel, if the analyzed color similarity index mai, texture similarity index mbi and shape similarity index mci are lower than the set color reference similarity index, texture reference similarity index and shape reference similarity index, it means that the similarity is higher than expected; extract the color similarity index mai, texture similarity index mbi and shape similarity index mci of the new product image and various image samples, and substitute them into the formula Perform weighted calculation to obtain the category bias index tri between the current new product and the various products in the classification model; , as well as They are the influencing weight factors of color similarity index mai, texture similarity index mbi and shape similarity index mci; It should be noted that, based on the above similarity analysis results of color, texture and shape features, we can comprehensively evaluate the similarity between the new product image and the features of various products in the classification model. According to the importance of different features, we assign corresponding weights to the similarity of each feature, and then take the weighted sum to get the final similarity measure.

[0019] Output commodity categories: According to the category bias index tri between the current new commodity and various commodities in the classification model, the category corresponding to the highest category bias index tri is selected from each group of category bias index tri as the estimated category of the current new commodity, and the new commodity is transported to the storage location corresponding to the estimated category in the commodity library through automated equipment; the automated equipment is specifically a sorting robot; It should be noted that by calculating the category bias index of new products and various categories of products in the classification model, the category corresponding to the highest index is selected as the estimated category. By using the classification model and quantitative indicators, a more accurate classification can be performed based on the image features of the new products, which helps the product library to manage the products in an orderly manner. At the same time, the second highest category bias index tri is selected, and the difference between it and the highest category bias index tri is calculated to obtain the category deviation value. If the calculated category deviation value is lower than the preset reference deviation value, the category corresponding to the second highest category bias index is used as the possible category, and the possible category and the estimated category are integrated and sent to the management personnel. The difference between the highest and second highest category bias indexes is calculated to obtain the category deviation value, and compared with the preset reference deviation value. If the deviation value is low, it means that the new product is not highly distinguished between the two categories, and there is a risk of inaccurate classification. This potential risk is notified to the management personnel in the form of possible categories and estimated categories, which plays a role in risk warning.

[0020] Product processing optimization: The constructed classification model is deployed to each node for parallel processing. When a new product needs to enter the product library, the load parameters of each node are analyzed to obtain the load evaluation index fzp of each node, and the node with the lowest load evaluation index fzp is selected as the distribution node for the new product image; the load parameters include task queue length, average task processing time, CPU usage rate, and memory usage rate; Specifically: Get the number of new product images waiting to be processed at each node at the current time point, so as to obtain the task queue length of each node; A long task queue will increase the processing delay of new tasks and affect the response speed of the entire system. For example, if a node has 50 pending product image tasks in its task queue, the newly added task may need to wait for the previous tasks to be processed in sequence. Obtain the processing time used by each node to process E groups of new product images before the current time point; where E>3, the specific value is set by the technical staff; thereby calculating the average task processing time of each node at the current time point, which is calculated by dividing the processing time used by E; The average task processing time reflects the actual processing capacity and efficiency of the node. If the average task processing time is long, it means that the node may be under a high load and has limited ability to process new tasks. Get the CPU usage and memory usage of each node at the current time point, set the weight coefficients corresponding to the CPU usage and memory usage, multiply the CPU usage and memory usage of each node at the current time point by the corresponding set weight coefficients, and then sum them up to get the operation status index of each node at the current time point; Preset the length value range, duration value range and index value range of each group corresponding to the task queue length, average task processing time and operation status index respectively; set each group of length value range to correspond to a task load score, each group of duration value range to correspond to a node timeliness score, and each group of index value range to correspond to a node load score; Match the task queue length, average task processing time and operation status index of each node at the current time point with the corresponding length value range, time value range and index value range of each group, so as to determine the task load score, node timeliness score and node load score of each node at the current time point; The task load score, node time efficiency score and node load score range are all set to 1-10. The shorter the task queue length, the higher the corresponding task load score; the shorter the average task processing time, the higher the corresponding node time efficiency score; the lower the operation status index, the higher the corresponding node load score; Additional notes, such as; Task queue length: Range I: 0-5, Task load score: 10 points; Range II: 6-15, Task load score: 8 points; Range III: 16-30, Task load score: 6 points; Range IV: 31-50, Task load score: 4 points; Range V: greater than 50, task load score: 2 points; Average task processing time: Range VI: 0-3 seconds, node timeliness score: 10 points; Range VII: 4-6 seconds, node timeliness score: 8 points; Range VIII: 7-10 seconds, node timeliness score: 6 points; Range IX: 11-15 seconds, node timeliness score: 4 points; Range X: greater than 15 seconds, node timeliness score: 2 points; Operational status index: Range Ⅺ: 0-0.3, node load score: 10 points; Range Ⅻ: 0.31-0.5, node load score: 8 points; Range XIII: 0.51-0.7, node load score: 6 points; Range XIV: 0.71-0.9, node load score: 4 points; Range XV: greater than 0.9, node load score: 2 points; The scoring of each node is determined by: Node A Task queue length: Assumed to be 12, falling in range II, task load score = 8 points; Average task processing time: Assumed to be 7 seconds, falling in range VIII, node timeliness score = 6 points; Operational status index: The calculated result is 0.44, which falls in range Ⅻ, and the node load score = 8 points; Node B Task queue length: Assumed to be 35, falling in range IV, task load score = 4 points; Average task processing time: Assumed to be 12 seconds, falling in range IX, node timeliness score = 4 points; Operational status index: The calculated result is 0.64, which falls in range XIII, and the node load score = 6 points; Node C Task queue length: Assumed to be 3, falling in range Ⅰ, task load score = 10 points; Average task processing time: Assumed to be 5 seconds, falling into range VII, node timeliness score = 8 points; Operational status index: The calculated result is 0.24, which falls in range Ⅺ, and the node load score = 10 points; Set the weight coefficients corresponding to the task load score, node timeliness score and node load score respectively, multiply the task load score, node timeliness score and node load score of each node at the current time point by the corresponding set weight coefficients, and then sum them up to obtain the load evaluation index fzp of each node at the current time point; It should be noted that by comprehensively considering multiple load parameters such as task queue length, average task processing time, CPU utilization and memory utilization, the load assessment index of each node is calculated, and the new product image is assigned to the node with the lowest load assessment index. This precise task allocation method can avoid assigning new tasks to nodes that are already in a high-load state, ensuring that tasks can be processed as soon as possible, thereby improving image processing efficiency.

[0021] Classification model verification: After the classification model corresponding to the current product library is established, according to the set evaluation time interval, after the set evaluation time interval is reached, X product images are randomly extracted from each category of product images in the pre-built image database; where X>5, the specific value is set by the technician; and input into the classification model for verification, so as to obtain the classification effect evaluation index yu of the classification model at the current time point, and according to the classification effect evaluation index yu of the classification model at the current time point, selectively trigger the adjustment signaling and send it to the technician; The pictures in the picture database are dynamically updated according to a preset update time interval; Specifically: Input the extracted X product images of various categories into the established classification model, and output the estimated categories corresponding to the X product images; The estimated categories corresponding to the X product images of each category are matched with the extracted category areas of the image database. If the match is successful, it is judged as a correct classification, otherwise it is judged as an abnormal classification; For the X product images extracted from each category, the percentage of correct classifications in each category is counted, that is, calculated by the number of correct classifications / X, and the classification accuracy rate fai of the classification model for each category of products is obtained; Preset the passing accuracy rate of various types of goods and mark them as fbi; According to the formula The classification accuracy fai of various types of commodities is weighted to obtain the classification effect evaluation index yu of the classification model at the current time point; is the influencing weight factor of the classification accuracy fai of various types of goods; Compare the classification effect evaluation index yu of the classification model at the current time point with the preset effect threshold index. If the classification effect evaluation index yu of the classification model at the current time point is less than the preset effect threshold index, trigger an adjustment signaling and send it to the technician; It should be noted that the classification effect evaluation index is obtained by weighted calculation based on the influencing weight factors of the classification accuracy of each category of goods; the importance differences of different categories of goods are taken into consideration and weight distribution is used to make the evaluation results more in line with actual business needs and more comprehensive and accurate reflection of the comprehensive performance of the model. Through continuous monitoring and timely feedback on model performance, it is ensured that the goods entering the product library can be accurately classified. Accurate classification helps in the storage, retrieval and management of goods, improves the overall efficiency and quality of product library operations, and reduces management costs and business risks caused by classification errors.

[0022] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for processing product appearance images for a large-scale product library, characterized in that: include: Establish a classification model: collect image samples of various categories of goods stored in the product library, and pre-process the image samples of various categories of goods; After the preprocessing is completed, feature extraction is performed on the preprocessed image samples of various categories of commodities, and the color features, texture features and shape features of the image samples of various categories of commodities stored in the commodity library are extracted. According to the pre-selected model architecture, the extracted features are input into the model structure to establish a classification model corresponding to the current commodity library; Input product image: When a new product needs to enter the product library, obtain the new product image and extract color features, texture features and shape features after preprocessing, and input it into the classification model for comprehensive evaluation, so as to output the category bias index tri between the current new product and each category of products in the classification model; where i represents the number of each category of products, i=1, 2, ..., k, k is the total number of product categories; Output product category: According to the category bias index tri between the current new product and each category of products in the classification model, select the category corresponding to the highest category bias index tri from each group of category bias index tri as the estimated category of the current new product, and transport the new product to the storage location of the corresponding estimated category in the product library through automated equipment.

2. The method for processing product appearance images for a large-scale product library according to claim 1, characterized in that: Obtain new product images and perform preprocessing to extract color features, texture features, and shape features, and input them into the classification model for comprehensive evaluation, specifically: S1: For the color features of the new product image and the samples of each type of product image in the classification model, count the number of pixels in each color interval of the new product image and the samples of each type of product image to obtain the color histogram of the new product image and the samples of each type of product image; Calculate the Bhattacharyya distance between the new product image and the color histogram of each type of product image sample as the color similarity index mai between each type of product image sample and the new product image; S2: Generate gray-level co-occurrence matrices for the texture features of the new product image and the various product image samples in the classification model; extract contrast, correlation, energy, and entropy features from the gray-level co-occurrence matrix, and thereby construct texture feature vectors for the new product image and the various product image samples in the classification model; Calculate the Mahalanobis distance between the texture feature vectors of the new product image and the samples of various product images as the texture similarity index mbi between the samples of various types of images and the new product image; S3: For the shape features of the new product image and the various product image samples in the classification model, extract the contours of the new product image and the various product image samples, and calculate the perimeter, area, circularity and eccentricity of the contours, thereby constructing shape descriptors of the new product image and the various product image samples in the classification model; The Euclidean distance is used to calculate the distance between the new product image and the shape descriptors of various product images, which is used as the shape similarity index mci between various image samples and the new product image.

3. The method for processing product appearance images for a large-scale product library according to claim 2, characterized in that: Output the category bias index tri between the current new product and each category of products in the classification model, specifically: S4: preset the color reference similarity index, texture reference similarity index and shape reference similarity index, and mark them as yai, ybi and yci; extract the color similarity index mai, texture similarity index mbi and shape similarity index mci of the new product image and various image samples, and substitute them into the formula Perform weighted calculation to obtain the category bias index tri between the current new product and the various products in the classification model; , as well as They are the influence weight factors of color similarity index mai, texture similarity index mbi and shape similarity index mci respectively.

4. The method for processing product appearance images for a large-scale product library according to claim 3, characterized in that: Also includes: Product processing optimization: The constructed classification model is deployed on each node for parallel processing. When a new product needs to enter the product library, the load parameters of each node are analyzed to obtain the load evaluation index fzp of each node, and the node with the lowest load evaluation index fzp is selected as the distribution node for the new product image; The load parameters include task queue length, average task processing time, CPU usage, and memory usage.

5. The method for processing product appearance images for a large-scale product library according to claim 4, characterized in that: Analyze the load parameters of each node, specifically: Get the number of new product images waiting to be processed at each node at the current time point, so as to obtain the task queue length of each node; Obtain the processing time used by each node to process E groups of new product images before the current time point; thereby calculate the average task processing time of each node at the current time point, which is calculated by dividing the processing time used by E; Get the CPU usage and memory usage of each node at the current time point, set the weight coefficients corresponding to the CPU usage and memory usage, multiply the CPU usage and memory usage of each node at the current time point by the corresponding set weight coefficients, and then sum them up to get the computing status index of each node at the current time point.

6. The method for processing product appearance images for a large-scale product library according to claim 5, characterized in that: Get the load evaluation index fzp of each node, specifically: Preset the length value range, duration value range and index value range of each group corresponding to the task queue length, average task processing time and operation status index respectively; set each group of length value range to correspond to a task load score, each group of duration value range to correspond to a node timeliness score, and each group of index value range to correspond to a node load score; Match the task queue length, average task processing time and operation status index of each node at the current time point with the corresponding length value range, time value range and index value range of each group, so as to determine the task load score, node timeliness score and node load score of each node at the current time point; Set the weight coefficients corresponding to the task load score, node timeliness score and node load score respectively, multiply the task load score, node timeliness score and node load score of each node at the current time point by the corresponding set weight coefficients, and then sum them up to obtain the load evaluation index fzp of each node at the current time point.

7. The method for processing product appearance images for a large-scale product library according to claim 6, characterized in that: Also includes: Classification model verification: After the classification model corresponding to the current product library is established, according to the set evaluation time interval, after the set evaluation time interval is reached, X product images are randomly extracted from each type of product image in the pre-built image database; and input into the classification model for verification, so as to obtain the classification effect evaluation index yu of the classification model at the current time point, and according to the classification effect evaluation index yu of the classification model at the current time point, selectively trigger the adjustment signaling and send it to the technical staff.

8. The method for processing product appearance images for a large-scale product library according to claim 7, characterized in that: According to the classification effect evaluation index yu of the classification model at the current time point, the selective trigger adjustment signaling is sent to the technician, specifically: Input the extracted X product images of various categories into the established classification model, and output the estimated categories corresponding to the X product images; The estimated categories corresponding to the X product images of each category are matched with the extracted category areas of the image database. If the match is successful, it is judged as a correct classification, otherwise it is judged as an abnormal classification; For the X product images extracted from each category, the percentage of correct classifications in each category is counted, that is, calculated by the number of correct classifications / X, and the classification accuracy rate fai of the classification model for each category of products is obtained; Preset the passing accuracy rate of various types of goods and mark them as fbi; According to the formula The classification accuracy fai of various types of commodities is weighted to obtain the classification effect evaluation index yu of the classification model at the current time point; is the influencing weight factor of the classification accuracy fai of various types of goods; The classification effect evaluation index yu of the classification model at the current time point is compared with the preset effect threshold index. If the classification effect evaluation index yu of the classification model at the current time point is less than the preset effect threshold index, the adjustment signal is triggered and sent to the technician.

Images