A shoebox target recognition method for intelligent picking

Abstract

The present invention relates to the field of image processing technology, and in particular to a shoebox target recognition method for intelligent picking, comprising: setting a reference block and multiple neighborhood blocks for each pixel point in a grayscale image of a shoebox, taking the normalized value of the product of the contrast of the pixel points of the reference block and the information entropy of the pixel points as a texture feature factor, determining the importance of the reference block according to the number and gradient value of the pixel points with gradient changes in the reference block, and the distance between the central pixel point of the reference block and the edge of the grayscale image, comprehensively quantizing the feature value of the reference block in combination with the importance and the texture feature factor, determining the similarity between the reference block and the neighborhood block according to the difference between the feature values ​​of the two, performing non-local mean filtering on the grayscale image based on the similarity, performing target detection and recognition on the filtered grayscale image, so as to complete intelligent picking, thereby improving the filtering effect and the accuracy of subsequent target detection.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and more particularly to a shoebox target recognition method for intelligent picking. Background Art

[0002] The intelligent picking system integrates advanced image recognition technology and target detection technology, which can accurately identify and classify shoe boxes, reduce the time and labor intensity of manual operation, and reduce human errors, ensuring that each shoe box can be correctly picked and processed, which is conducive to refined inventory management. During the transmission of shoe box images, the collected shoe box images may produce salt and pepper noise due to changes in light or electromagnetic interference in the environment. The presence of these noise pixels may cause the detection algorithm to mistakenly identify the noise points as part of the shoe box, which in turn causes misclassification and reduces the accuracy of shoe box target recognition. Therefore, before target detection, the shoe box image needs to be filtered and denoised to improve image quality and ensure the accuracy of subsequent classification.

[0003] For the filtering and denoising of images, the existing Chinese patent document with the publication number CN109064418B discloses a non-uniform noise image denoising method based on non-local mean, including: Step 1: Using the evaluation operator First, the non-uniform noise image is roughly classified by pixels; Step 2: For each pixel in the noise image, according to the rough classification results of its surrounding neighborhood pixels, the majority voting method is used to finely classify the category of the pixel into one of the following categories: low noise high texture, medium texture, high noise sub-texture, and smooth area; Step 3: For each category after fine classification, the filter parameters and neighborhood block size are adaptively selected, and the non-local mean denoising algorithm is used to perform pixel denoising.

[0004] The non-local mean filtering algorithm in the above scheme is a widely used technology for image denoising. Its basic idea is to use the information of similar areas in the image to remove noise. Specifically, the algorithm will find some other areas (neighborhood blocks) similar to the surrounding areas (reference blocks) for each pixel, and then update the pixel value of the pixel based on the information of these similar areas. When calculating the similarity between the reference block and the neighborhood block of each pixel, the grayscale difference is usually used.

[0005] However, when there are noisy pixels in the image, these noise points will cause the grayscale difference between the reference block and the neighboring block to increase significantly. For example, an originally smooth area becomes irregular due to the presence of noise points, which makes the blocks that should be considered similar be mistakenly judged as dissimilar. This deviation leads to errors in similarity calculation, which in turn affects the filtering effect, resulting in poor denoising results and ultimately affecting the accuracy of subsequent classification and recognition of shoe boxes. Summary of the invention

[0006] In order to solve the problem that when performing non-local mean filtering denoising on a shoebox image, the similarity between the reference block and the neighborhood block is judged only based on the grayscale difference, resulting in errors in similarity calculation, affecting the filtering effect and the accuracy of subsequent classification and recognition of the shoebox, the present invention proposes a shoebox target recognition method for intelligent picking, comprising:

[0007] A reference block and a plurality of neighborhood blocks are set for each pixel in the grayscale image of the shoe box, and a normalized value of the product of the contrast of the pixel of the reference block and the information entropy of the pixel is used as a texture feature factor of the reference block;

[0008] Determine pixels with gradient changes in the reference block, and determine the importance of the reference block according to the ratio of the number of pixels with gradient changes to the mode of the gradient value and the distance between the central pixel of the reference block and the edge of the grayscale image, wherein the importance is positively correlated with the ratio and negatively correlated with the distance;

[0009] Comprehensively calculating a feature value of the reference block in combination with the importance and the texture feature factor, wherein the feature value is positively correlated with both the importance and the texture feature factor;

[0010] Determine the similarity between the reference block and each of the neighborhood blocks according to the difference between the feature value of the reference block and the feature value of each of the neighborhood blocks; the method for determining the feature value of each of the neighborhood blocks is the same as the method for determining the feature value of the reference block;

[0011] Non-local mean filtering is performed on the grayscale image based on the similarity between the reference block and the neighborhood block of all pixels, and target detection and recognition is performed on the filtered grayscale image using a target detection network.

[0012] The reference block set in the above technical solution is used for feature extraction of the current pixel, while the neighborhood block is used to find similar areas, which provides a basis for subsequent texture feature extraction and similarity calculation. Contrast reflects the brightness change of the image, and information entropy reflects the information richness of the image. The texture feature factor combines these two indicators, which can more accurately describe the texture features of the reference block, thereby improving the accuracy of similarity calculation. The importance of the reference block is determined by gradient change and position information. The area with more gradient changes usually contains more details and edge information, and the area close to the edge of the image is more likely to be a shoebox label, and the higher the importance. In this way, the importance of the reference block in the image can be better evaluated, so that more attention is paid to key areas during the filtering process. The importance and texture feature factor are combined to form a comprehensive feature value, which can more comprehensively reflect the characteristics and importance of the reference block. The similarity between the reference block and the neighborhood block is determined by comparing their feature values. This method takes into account texture features and importance, can more accurately find similar areas, improve the accuracy of similarity calculation, and reduce the influence of noise on similarity calculation, thereby improving the filtering effect and the accuracy of subsequent classification and recognition of shoebox images.

[0013] Preferably, the importance of the reference block satisfies the following relationship:

[0014]

[0015] In the formula, For the The importance of the reference block of pixels, For the The number of pixels with gradient changes in the reference block of pixels. For the The mode of the gradient values ​​of the pixels with gradient changes in the reference block of pixels. is the natural exponential function, For the The Euclidean distance between the center pixel of the reference block with pixels and the closer edge pixel of the two horizontal edge pixels of the grayscale image is For the The Euclidean distance between the central pixel of a reference block of pixels and the closer one of the two edge pixels in the vertical direction of the grayscale image.

[0016] The above technical solution can more accurately evaluate the importance of each reference block by analyzing the gradient changes of the pixel points of the reference block and the position of the reference block. The higher the importance, the more critical the information of the reference block in the image, thereby better retaining key information during filtering and target recognition.

[0017] Preferably, the characteristic value of the reference block satisfies the following relationship:

[0018]

[0019] In the formula, For the The feature value of the reference block of pixels, For the The texture feature factor of the reference block of pixels, For the The importance of the reference block of each pixel.

[0020] The above technical solution obtains a comprehensive eigenvalue by multiplying the texture feature factor and the importance. This eigenvalue takes into account both the texture complexity of the reference block and its importance in the image. This comprehensive quantification method can more comprehensively describe the characteristics of the reference block.

[0021] Preferably, the similarity between the reference block and each of the neighborhood blocks satisfies the following relationship:

[0022]

[0023] In the formula, For the The reference block of pixels is The pixel The similarity between neighboring blocks, For the The feature value of the reference block of pixels, For the The pixel The eigenvalues ​​of the neighborhood blocks, To take the absolute value sign.

[0024] The above technical solution uses the natural exponential function The difference in eigenvalues ​​is converted into a similarity between 0 and 1 to obtain a more accurate similarity calculation result. Based on the similarity, similar neighborhood blocks can be found more effectively, thereby improving the effect of non-local mean filtering.

[0025] Preferably, the method of setting the reference block and multiple neighborhood blocks for each pixel in the grayscale image of the shoe box is:

[0026] With each pixel as the center, define a size of A window, all pixels contained in the window constitute a reference block of the pixel;

[0027] Define the search range of the pixel as the entire grayscale image, traverse each pixel in the search range, take the pixel as the center, and the size is All pixels within the window are regarded as a neighborhood block;

[0028] in, is the default value.

[0029] Preferably, the method for determining the pixel points having gradient changes in the reference block is:

[0030] In the reference block, adjacent pixels in the upper, lower, left and right directions of each pixel are obtained. If the grayscale values ​​of the adjacent pixels are the same as the grayscale value of the pixel, the pixel is a pixel with no gradient change in the reference block; if the grayscale value of one or more of these adjacent pixels is different from the grayscale value of the pixel, the pixel is a pixel with gradient change in the reference block.

[0031] Preferably, the contrast of the pixels of the reference block is determined according to the standard deviation of the grayscale values ​​of all the pixels in the reference block.

[0032] The above technical solution reflects the degree of change of the grayscale values ​​of pixels in the reference block through the standard deviation of the grayscale values ​​of all pixels in the reference block, which can better capture the visual features in the reference block and improve the accuracy of feature extraction. The larger the standard deviation, the more drastic the change in the grayscale value, that is, the higher the contrast.

[0033] Preferably, the performing non-local mean filtering on the grayscale image based on the similarity between the reference block and the neighborhood block of all pixels comprises:

[0034] For each pixel reference block, select the pixel with the closest similarity to the reference block. Neighborhood blocks, is the total number of neighborhood blocks that are closest to the reference block in similarity;

[0035] The filtering operation is based on the following relationship:

[0036]

[0037] In the formula, For the The filtered gray value of each pixel is is the serial number of the neighboring block that is closest to the reference block in similarity, For the The reference block with the closest similarity to the The gray value of the central pixel of the neighborhood block is For the The reference block with the pixel point is closest to the The similarity between neighboring blocks.

[0038] Preferably, the method of using the target detection network to perform target detection and recognition on the filtered grayscale image is:

[0039] Collect multiple types of shoe box images in advance, mark the label area in each shoe box image, and assign a corresponding label to each shoe box category; the label area includes a brand logo and a model number;

[0040] The labeled shoebox images are used to train the object detection network. During the training process, the network learns how to recognize different types of shoeboxes and their label areas.

[0041] The filtered grayscale image is input as the input image to the trained object detection network, and the object detection network outputs the category probability of the shoebox in the input image and its corresponding confidence.

[0042] Set a confidence threshold and retain detection results with confidence higher than or equal to the threshold to finally determine the category of the shoebox in the input image.

[0043] The above technical solution improves the accuracy of target detection through high-quality data sets and effective training process, combined with confidence threshold filtering, thereby improving the efficiency and accuracy of shoe box classification and picking.

[0044] The present invention has the following effects:

[0045] The present invention obtains texture feature factors by calculating contrast and information entropy, and uses gradient changes and position information to determine the importance of reference blocks. The texture feature factors and importance are combined to form a comprehensive feature value, which significantly improves the accuracy and robustness of similarity calculation. The similarity is determined by comparing the comprehensive feature values ​​of the reference block and the neighborhood block, effectively reducing the impact of noise on the filtering effect and improving the quality of image denoising. Ultimately, this improved method improves the accuracy of shoebox image classification and recognition and ensures the reliability of the intelligent picking system. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0047] Figure 1 It is a schematic flow chart of the method of the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0049] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0050] Reference Figure 1 The present invention provides a shoebox target recognition method for intelligent picking, comprising steps S1 to S7:

[0051] S1: Set the reference block and multiple neighborhood blocks for each pixel in the grayscale image of the shoebox.

[0052] Since the present invention optimizes the non-local mean filtering algorithm to improve the filtering effect, obtains high-quality images, uses high-quality images to input the target detection network, and thus improves the accuracy of target detection and recognition, and the optimized non-local mean filtering algorithm performs filtering operations based on the similarity between the reference block and the neighborhood block of the pixel point, this step first sets the reference block and multiple neighborhood blocks of each pixel point in the grayscale image, and the specific method is:

[0053] For any pixel in the grayscale image, the reference block is a small area centered on the pixel. Specifically, a region with a size of A window is formed, and all pixels contained in the window constitute a reference block of the pixel, that is, the area corresponding to the window range is used as the reference block of the pixel.

[0054] For any pixel in the grayscale image, the neighborhood block is the area with the same size as the reference block, and the similarity between the neighborhood block and the reference block needs to be analyzed later. Therefore, the search range of the pixel is defined as the entire grayscale image, and each pixel in the search range is traversed, with the pixel as the center, and the size is All pixels in the window are regarded as a neighborhood block, and multiple neighborhood blocks with the same size as the reference block will be generated within the search range of the pixel.

[0055] in, is the default value.

[0056] In the non-local mean filtering algorithm, the selection of reference blocks and search range has an important influence on the filtering effect. For reference blocks, smaller reference blocks can better capture local details and reduce the amount of calculation, but are easily affected by noise. Larger reference blocks can contain more context information and improve the robustness of similarity calculation, but will increase the amount of calculation, usually The reasonable value range of is between 1 and 5. , is a more balanced choice that can achieve a good compromise between capturing details and reducing the impact of noise.

[0057] As for the search range, a smaller search range can reduce the amount of calculation and improve the processing speed, but it may not be able to find enough similar blocks, resulting in poor filtering effect. A larger search range can find more similar blocks and improve the filtering effect. Therefore, this step uses the entire grayscale image as the search range.

[0058] In summary, the size of the reference block set in this step is , the search range is the entire grayscale image.

[0059] S2: Calculate the texture feature factor of the reference block.

[0060] Since the traditional non-local mean filtering algorithm determines the similarity between the reference block and the neighborhood block by calculating the grayscale difference between them. However, when there are noisy pixels in the image, these noise points will significantly affect the calculation of the grayscale difference, resulting in inaccurate similarity calculation. Therefore, in order to improve the accuracy of the similarity calculation, it is necessary to find a method to better describe the texture features of each reference block, thereby reducing the impact of noise.

[0061] In this step, the texture feature factor of the reference block is calculated by the contrast of the pixel points of the reference block and the information entropy of the pixel points. The texture feature factor of the reference block is used to reflect the texture feature of the reference block. The calculation formula is:

[0062]

[0063] Formula explanation: Indicates The texture feature factor of the reference block of pixels, Indicates The contrast of the pixels in the reference block of pixels, Indicates The information entropy of the pixels of the reference block of pixels, is a normalization operation, so that The value of within the range.

[0064] In the formula, The larger the value is, the more obvious the difference in brightness of the pixels in the reference block is, which makes the texture clearer and more prominent. The greater the contrast of the pixels in the reference block, the denser and more detailed the texture will appear. The contrast of the pixels is characterized by the standard deviation of the grayscale values ​​of all pixels in the reference block, that is, The value of is equal to the standard deviation of the grayscale values ​​of all pixels in the reference block. The larger it is, the more complex the grayscale distribution in the area is, and there are more details and changes. Therefore, it usually means that the denser the texture, the larger the corresponding texture feature factor will be.

[0065] This formula calculates the texture feature factor of the reference block, which comprehensively considers the contrast of the pixels in the reference block and the information entropy of the pixels. These two indicators can more comprehensively reflect the texture characteristics of the image. Among them, the contrast of the pixels reflects the degree of change of the pixel values ​​in the reference block. The higher the contrast, the more obvious the difference in brightness of the pixels in the reference block, which can make the texture clearer and more prominent. Among them, information entropy is an indicator to measure the richness of information. It is used to describe the complexity of the distribution of pixel values ​​in image processing. In image processing, information entropy can be used to describe the complexity of the distribution of pixel values ​​in an image. Specifically, the higher the information entropy, the more uniform and complex the distribution of pixel values ​​in the image, and the richer the texture information in the image; the lower the information entropy, the more concentrated and simple the distribution of pixel values ​​in the image, and the less texture information in the image.

[0066] By calculating the texture feature factor of the reference block, the texture features of the reference block can be described more accurately, rather than relying solely on grayscale differences. This can reduce the impact of noise pixels on subsequent similarity calculations. The texture feature factor can better capture the complex texture information within the reference block, making it more accurate when searching for similar neighborhood blocks.

[0067] S3: Evaluate the importance of the reference block.

[0068] In image processing, if only the method in the above steps is used to calculate the texture feature factors of the reference block and the texture feature factors of the neighborhood block respectively, and the similarity between the reference block and the neighborhood block is measured based on the difference in texture feature factors, there are still limitations. Specifically, if the reference block and the neighborhood block may be very similar in texture, but they are actually located in different areas of the image and have different importance, only based on the difference in texture feature factors to measure the similarity between the reference block and the neighborhood block for non-local mean filtering will lead to inaccurate filtering results.

[0069] For example, the grayscale image corresponding to the shoebox image contains the following two areas:

[0070] A small piece of smooth background in the label area of ​​the shoebox (such as the blank part on the label) and a small piece of smooth background on the surface of the shoebox. There is no obvious texture change in these two images, and the texture features are relatively similar. In this case, if the similarity is calculated based on the texture features only and filtered based on the similarity, the two images will be processed in the same way during the filtering process. If both are filtered out, the key details (such as the smooth background of the label part) will be blurred or lost. If both are retained, unimportant details (such as the smooth background area on the surface of the shoebox) will be mistakenly retained, which will lead to inaccurate filtering results.

[0071] It can be seen that some areas (such as the shoe box label area) are very important in the image, but the importance of these areas cannot be distinguished based on the texture feature factor alone, which will affect the accuracy of the filtering results. Therefore, in order to more accurately determine the similarity between the reference block and the neighborhood block, this step needs to analyze the importance of the reference block (in the image).

[0072] Regarding the importance of the reference block, this step considers the following aspects:

[0073] Consistency of gradient value changes: The change in the gradient value of the pixel in the reference block reflects the change in the brightness or color of the pixel in the reference block. In the label area of ​​the shoebox, the change in the gradient value of the pixel is usually highly consistent. If the change in the gradient value of the pixel in the reference block is more consistent, it means that the reference block is more likely to be the label area of ​​the shoebox or other important information area.

[0074] Distance from the edge of the image: The label area is usually located at the edge or corner of the shoe box. These areas often contain special marks or sealing information. The closer the reference block is to the edge of the image, the more important the reference block is.

[0075] Therefore, the importance of the reference block satisfies the following relationship:

[0076]

[0077] In the formula, For the The importance of the reference block of pixels, For the The number of pixels with gradient changes in the reference block of pixels reflects the extent of the brightness or color changes of the pixels in the reference block. It is a symbol of wholeness. For the The mode of the gradient values ​​of the pixels with gradient changes in the reference block of pixels. This value reflects the magnitude of the gradient change of the pixels in the reference block. It is a symbol of wholeness. is the natural exponential function, For the The Euclidean distance between the central pixel of the reference block with 4 pixels and the closer edge pixel of the two edge pixels in the horizontal direction of the grayscale image is calculated. That is, first use edge detection to obtain all edge pixels of the grayscale image, obtain the two edge pixels that the central pixel of the reference block passes through in the horizontal direction, and calculate the Euclidean distance between the central pixel and the closer edge pixel. , It is a symbol of wholeness. For the The Euclidean distance between the central pixel of the reference block with pixels and the closer edge pixel of the two edge pixels in the vertical direction of the grayscale image is obtained, that is, the two edge pixels passed by the central pixel of the reference block in the vertical direction are obtained, and the Euclidean distance between the central pixel and the closer edge pixel is calculated as , It is also a symbol of an entirety.

[0078] In this formula, the method for obtaining the pixel points with gradient changes in the reference block is:

[0079] In the reference block, adjacent pixels in the upper, lower, left and right directions of each pixel are obtained. If the grayscale values ​​of the adjacent pixels are the same as the grayscale value of the pixel, the pixel is a pixel with no gradient change in the reference block; if the grayscale value of one or more of these adjacent pixels is different from the grayscale value of the pixel, the pixel is a pixel with gradient change in the reference block.

[0080] In this formula, This ratio reflects the consistency of the gradient changes of the pixels of the reference block. If the gradient changes of the pixels of the reference block are more consistent (that is, the larger the mode of the gradient value and the more pixels with gradient changes), the higher the ratio is, which means that the reference block may be the label area or other important information area of ​​the shoe box, and the reference block is more important. If the gradient changes of the pixels of the reference block are more inconsistent, the lower the ratio is, which means that the reference block may be the background of the shoe box or other non-important areas, and the reference block is less important.

[0081] In this formula, This index item reflects the relationship between the distance (between the reference block and the edge of the grayscale image) and the importance of the reference block. If the reference block is close to the edge of the image, it means that this area may be a label area or other important information area. The higher the value of this index item, the more important the reference block is. If the reference block is far away from the edge of the image, it means that this area may be the background or other unimportant areas. The lower the value of this index item, the less important the reference block is.

[0082] Therefore, this step comprehensively considers the consistency of gradient changes and the distance between the reference block and the edge of the image to more accurately evaluate the importance of each reference block in the image. By combining these two factors, important information areas (such as label areas) and non-important areas (such as background) can be more accurately distinguished, which helps to reduce the impact of noise during the filtering process while better preserving important details.

[0083] S4: The feature value of the reference block is comprehensively quantified by combining the importance of the reference block and the texture feature factor of the reference block.

[0084] This step comprehensively considers the texture feature factor of the reference block and the importance of the reference block (in the image) to quantify the feature value of the reference block, where the feature value of the reference block satisfies the following relationship:

[0085]

[0086] In the formula, For the The feature value of the reference block of pixels, For the The texture feature factor of the reference block of pixels, For the The importance of the reference block of each pixel.

[0087] In this formula, the texture feature factor is a value that reflects the texture complexity of the reference block. The higher the value, the more complex the texture. The importance reflects whether the reference block is located in an important information area. Combining the texture feature factor and the importance, we get a feature value of the reference block. This feature value can more comprehensively describe the texture features and importance of the reference block.

[0088] Based on this eigenvalue, similar neighborhood blocks can be found more accurately for the reference block, which can better guide the filtering process, making the filtering results more accurate and reliable, improving the filtering effect, and then in the subsequent shoebox target recognition task, more accurate focus can be placed on the label area and other important information areas, which can improve the accuracy of target recognition.

[0089] S5: Obtain the similarity between the reference block and the neighboring block according to the difference between the feature value of the reference block and the feature value of the neighboring block.

[0090] The similarity between the reference block and each of the neighborhood blocks satisfies the following relationship:

[0091]

[0092] In the formula, For the The reference block of pixels is The pixel The similarity between neighboring blocks, For the The feature value of the reference block of pixels, For the The pixel The eigenvalues ​​of the neighborhood blocks, To take the absolute value sign.

[0093] Among them, the method for determining the eigenvalue of each neighborhood block is the same as the method for determining the eigenvalue of the reference block, that is, first calculate the texture feature factor of the neighborhood block according to the method of step S2, evaluate the importance of the neighborhood block according to the method of step S3, and obtain the eigenvalue of the neighborhood block according to the method of step S4.

[0094] The eigenvalue calculated by this formula comprehensively considers the texture features and importance. By comparing the difference between the eigenvalues ​​of the reference block and the neighborhood block, the similarity between the reference block and the neighborhood block is determined, and similar neighborhood blocks can be found more accurately. If the difference between the eigenvalues ​​is smaller, the similarity between the reference block and the neighborhood block is higher, which means that the reference block and the neighborhood block are more similar in texture features and importance. In this case, the neighborhood block has a greater impact on the filtering result of the reference block, and vice versa.

[0095] In the filtering process, more consideration is given to neighborhood blocks with high similarity, which can better preserve important details while reducing the impact of noise, thereby improving the accuracy and reliability of the filtering results. Neighborhood blocks with high similarity usually contain more useful information, while neighborhood blocks with low similarity may contain noise or irrelevant information.

[0096] S6: Perform non-local mean filtering based on the similarity between the reference block and the neighboring blocks.

[0097] For each pixel reference block, select the pixel with the closest similarity to the reference block. Neighborhood blocks, is the total number of neighborhood blocks that are closest to the reference block in similarity. , this value is an experience value.

[0098] Next, the filtering operation is performed based on the following relationship:

[0099]

[0100] In the formula, For the The filtered gray value of each pixel is is the serial number of the neighboring block that is closest to the reference block in similarity, It is a symbol of wholeness. For the The reference block with the closest similarity to the The gray value of the central pixel of the neighborhood block is For the The reference block with the closest similarity The similarity between neighboring blocks.

[0101] The formula is As The weights of the neighborhood blocks with high similarity are used to fuse the grayscale values ​​of the neighborhood blocks with high similarity by weighted averaging to obtain the grayscale value after filtering. This can effectively remove noise while retaining important details of the image and improving the image quality. This method is particularly suitable for processing images containing important texture and structural information, such as the label area of ​​a shoe box. Similarity plays a key role here, which determines the contribution of each neighborhood block to the filtering result. The higher the similarity, the greater the contribution.

[0102] S7: Use the target detection network to perform target detection and recognition on the filtered grayscale image.

[0103] In automated warehouses or logistics centers, quickly and accurately identifying information on shoe boxes is crucial to improving the efficiency of the picking system. The object detection network can accurately locate key areas in shoe box images, such as brand labels, model information, and barcodes / QR codes, and can also classify detected objects, thereby improving the accuracy and efficiency of intelligent picking systems.

[0104] The specific process of target detection and recognition in this step is as follows:

[0105] Data collection and labeling: Collect various types of shoebox images in advance, mark the label area (such as brand logo, model, etc.) in each shoebox image, and assign corresponding labels to each shoebox category.

[0106] Model training: Use the labeled shoebox images to train the target detection network. The target detection network selects the CNN (convolutional neural network) network and sets the classification loss function to the cross entropy loss function or the binary cross entropy loss function. These two functions are the most commonly used and effective classification loss functions. The classification loss function is a function used in machine learning and deep learning to measure the difference between the model prediction and the true label. During the training process, the CNN network will learn how to identify different types of shoeboxes and their label areas.

[0107] Target detection: The grayscale image of the shoebox to be identified after filtering is input into the trained target detection network. The target detection network outputs the probability that the shoebox to be identified belongs to each shoebox category and the confidence level corresponding to each probability.

[0108] If the confidence threshold is set to 0.8 (empirical value), then only when the confidence that the shoebox to be detected belongs to a certain category is greater than or equal to 0.8, the model's prediction of the category is considered reliable. The maximum probability value is found in the detection results with confidence greater than or equal to 0.8, and the detection result corresponding to the maximum probability value is used as the final classification recognition result. That is, if the probability that the shoebox to be identified belongs to a certain shoebox category is the largest, and the confidence of the probability is greater than or equal to the confidence threshold, the shoebox to be identified is classified as the shoebox category. For example, assuming that there are shoebox categories A, B, and C, the target detection network output results are: the probability that the shoebox to be identified belongs to A is 0.2, and the confidence of the probability is 0.7; the probability that the shoebox to be identified belongs to B is 0.1, and the confidence of the probability is 0.8; the probability that the shoebox to be identified belongs to C is 0.7, and the confidence of the probability is 0.9. Then the shoebox to be identified is classified as C.

[0109] Through the above steps, efficient and accurate target detection and classification of the filtered shoe box grayscale image can be achieved, thereby improving the efficiency and accuracy of the intelligent picking system.

[0110] Combining steps S1 to S7, this solution redefines the similarity measurement method between the reference block and the neighborhood block by analyzing the texture features of the reference block corresponding to each pixel point that needs to be filtered and its importance in the image. The new similarity measurement method can more accurately reflect the actual similarity between the two blocks, reduce the impact of noise and retain the real details of the image, and can achieve more accurate filtering of the shoebox image and shoebox target recognition, thereby improving the overall performance and accuracy of the intelligent picking system.

[0111] In the description of this specification, "plurality" or "several" means at least two, such as two, three or more, etc., unless otherwise clearly and specifically defined.

[0112] Although this specification has shown and described a number of embodiments of the present invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, changes and alternatives without departing from the ideas and spirit of the present invention. It should be understood that in the practice of the present invention, various alternatives to the embodiments of the present invention described herein may be employed.

Claims

1. A shoebox target recognition method for intelligent picking, characterized in that: include: A reference block and a plurality of neighborhood blocks are set for each pixel in the grayscale image of the shoe box, and a normalized value of the product of the contrast of the pixel of the reference block and the information entropy of the pixel is used as a texture feature factor of the reference block; Determine pixels with gradient changes in the reference block, and determine the importance of the reference block according to the ratio of the number of pixels with gradient changes to the mode of the gradient value and the distance between the central pixel of the reference block and the edge of the grayscale image, wherein the importance is positively correlated with the ratio and negatively correlated with the distance; The characteristic value of the reference block is calculated comprehensively by combining the importance and the texture characteristic factor, satisfying the following relationship: ; In the formula, For the The feature value of the reference block of pixels, For the The texture feature factor of the reference block of pixels, For the The importance of the reference block of the pixel points; the eigenvalue is positively correlated with the importance and the texture feature factor; Determine the similarity between the reference block and each of the neighborhood blocks according to the difference between the feature value of the reference block and the feature value of each of the neighborhood blocks; the method for determining the feature value of each of the neighborhood blocks is the same as the method for determining the feature value of the reference block; Non-local mean filtering is performed on the grayscale image based on the similarity between the reference block and the neighborhood block of all pixels, and target detection and recognition is performed on the filtered grayscale image using a target detection network.

2. The shoebox target recognition method for intelligent picking according to claim 1 is characterized in that: The importance of the reference block satisfies the following relationship: ； In the formula, For the The importance of the reference block of pixels, For the The number of pixels with gradient changes in the reference block of pixels. For the The mode of the gradient values ​​of the pixels with gradient changes in the reference block of pixels. is the natural exponential function, For the The Euclidean distance between the center pixel of the reference block with pixels and the closer edge pixel of the two horizontal edge pixels of the grayscale image is For the The Euclidean distance between the central pixel of a reference block of pixels and the closer one of the two edge pixels in the vertical direction of the grayscale image.

3. The shoebox target recognition method for intelligent picking according to claim 1 is characterized in that: The similarity between the reference block and each of the neighborhood blocks satisfies the following relationship: ； In the formula, For the The reference block of pixels is The pixel The similarity between neighboring blocks, For the The feature value of the reference block of pixels, For the The pixel The eigenvalues ​​of the neighborhood blocks, To take the absolute value sign.

4. The shoebox target recognition method for intelligent picking according to claim 1 is characterized in that: The method of setting the reference block and multiple neighborhood blocks for each pixel in the grayscale image of the shoebox is: With each pixel as the center, define a size of A window, all pixels contained in the window constitute a reference block of the pixel; Define the search range of the pixel as the entire grayscale image, traverse each pixel in the search range, take the pixel as the center, and the size is All pixels within the window are regarded as a neighborhood block; in, is the default value.

5. The shoebox target recognition method for intelligent picking according to claim 1 is characterized in that: The method for determining the pixel points with gradient changes in the reference block is: In the reference block, adjacent pixels in the upper, lower, left and right directions of each pixel are obtained. If the grayscale values ​​of the adjacent pixels are the same as the grayscale value of the pixel, the pixel is a pixel with no gradient change in the reference block; if the grayscale value of one or more of these adjacent pixels is different from the grayscale value of the pixel, the pixel is a pixel with gradient change in the reference block.

6. The shoebox target recognition method for intelligent picking according to claim 1 is characterized in that: The contrast of the pixels of the reference block is determined according to the standard deviation of the grayscale values ​​of all the pixels in the reference block.

7. The shoebox target recognition method for intelligent picking according to claim 1 is characterized in that: The performing non-local mean filtering on the grayscale image based on the similarity between the reference block and the neighborhood block of all pixels comprises: For each pixel reference block, select the pixel with the closest similarity to the reference block. Neighborhood blocks, is the total number of neighborhood blocks that are closest to the reference block in similarity; The filtering operation is based on the following relationship: ； In the formula, For the The filtered gray value of each pixel is is the serial number of the neighboring block that is closest to the reference block in similarity, For the The reference block with the closest similarity to the The gray value of the central pixel of the neighborhood block is For the The reference block with the pixel point is closest to the The similarity between neighboring blocks.

8. The shoebox target recognition method for intelligent picking according to claim 1, characterized in that: The method of using the target detection network to perform target detection and recognition on the filtered grayscale image is: Collect multiple types of shoe box images in advance, mark the label area in each shoe box image, and assign a corresponding label to each shoe box category; the label area includes a brand logo and a model number; The labeled shoebox images are used to train the object detection network. During the training process, the object detection network learns to recognize different types of shoeboxes and their label areas. The filtered grayscale image is input as the input image to the trained object detection network, and the object detection network outputs the category probability of the shoebox in the input image and its corresponding confidence. Set a confidence threshold and retain detection results with confidence higher than or equal to the threshold to finally determine the category of the shoebox in the input image.

Images