Warehouse logistics intelligent management method and system based on image classification

Abstract

The application relates to the technical field of image recognition, in particular to a warehouse logistics intelligent management method and system based on image classification. The method comprises the following steps: acquiring an image of a to-be-sorted cargo on a conveying device, and pre-processing the image to obtain a standard image; an adaptive diffusion factor of the standard image is constructed; the iteration number and time step of the AKAZE algorithm are adjusted through the adaptive diffusion factor, and feature points in the standard image are extracted and cargo classification is performed. The method can adaptively control the algorithm complexity, thereby improving the sorting speed.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, and in particular to a method and system for intelligent management of warehousing and logistics based on image classification. Background Technology

[0002] Intelligent warehousing and logistics systems primarily utilize computer vision technology for automated identification, sorting, and tracking of goods. Among these, image-based intelligent management technology can determine the type and status of goods directly by collecting the texture and shape features of their appearance, without relying on specific label media such as barcodes. This enhances the flexibility and robustness of the logistics system and is used to achieve automatic identification and sorting in unmanned warehousing scenarios.

[0003] In related technologies, images of goods on conveyor belts are often captured using industrial cameras, and the appearance features of the goods are modeled based on local feature extraction and matching algorithms. Among them, the AKAZE (Accelerated KAZE) algorithm, due to its use of nonlinear diffusion to construct the scale space, can better preserve image edge information during feature point localization and has a certain degree of rotation invariance and illumination adaptability, has been applied to scenarios such as target recognition and image classification.

[0004] However, warehousing and logistics scenarios have characteristics that differ from static shooting environments. For example, in warehousing and logistics scenarios, goods usually move with conveyor belts or automated equipment, which can easily cause motion blur during imaging. At the same time, common packaging such as cardboard boxes and plastic boxes have high similarity in texture and color, and the background of the logistics site is complex.

[0005] Under the aforementioned conditions, directly applying the original AKAZE algorithm results in extremely high computational complexity due to the heavy workload of pixel-by-pixel iterative solving required to construct the nonlinear scale space, and the lack of an adaptive adjustment mechanism for image quality (such as blur level or information density). When processing massive amounts of logistics images with fluctuating quality, the algorithm's feature extraction speed is slow, affecting the sorting of goods on the production line and leading to low sorting efficiency during production. Summary of the Invention

[0006] To address the problem of low efficiency in image feature recognition and low sorting speed caused by traditional image classification algorithms, this application provides an intelligent management method and system for warehousing and logistics based on image classification.

[0007] Firstly, this application provides an intelligent management method for warehousing and logistics based on image classification, employing the following technical solution: The image classification-based intelligent management method for warehousing and logistics includes: acquiring images of goods to be sorted on conveyor equipment and preprocessing the images to obtain standard images; The standard image is divided into multiple local windows; a gradient structure tensor matrix is ​​constructed for each local window, and texture blur is constructed based on the difference between the maximum and minimum eigenvalues ​​of the gradient structure tensor matrix to characterize the image blur. Local windows with information entropy values ​​greater than a preset entropy threshold are selected as reference tiles. Effective feature density is calculated based on the texture fuzziness and information entropy of the reference tiles. Effective feature density is negatively correlated with texture fuzziness and positively correlated with information entropy. Based on the distribution characteristics of the reference patches, the spatial distribution dispersion of the reference patches is calculated; the ratio of the width to the height of the standard image is used as the shape correction coefficient, and the product of the effective feature density, the reciprocal of the sum of the texture blur of each reference patch, the spatial distribution dispersion, and the shape correction coefficient is used as the adaptive diffusion factor. The number of iterations and time step of the AKAZE algorithm are adjusted by an adaptive diffusion factor, and feature points in standard images are extracted for cargo classification.

[0008] Texture blur is constructed through local window partitioning and gradient structure tensor analysis to characterize the directionality and degree of motion blur. Information entropy is then used to filter high-information regions, avoiding the consumption of computational resources by low-value regions. Furthermore, spatial distribution dispersion is used to characterize the breadth of effective feature distribution in the image, preventing misjudgment of overall image quality based solely on locally clear areas. Simultaneously, a shape correction coefficient is introduced to adapt the diffusion adjustment mechanism to the elongated images formed by linear scan cameras. Based on this, an adaptive diffusion factor directly participates in the parameter control of the AKAZE scale spatial construction process, enabling the algorithm to proactively reduce computational intensity when the image is blurry and information is sparse, and to fully extract details when the image is clear and features are rich. This significantly reduces the overall computational load while ensuring classification accuracy, thereby improving the throughput of the sorting system.

[0009] Optionally, the step of constructing texture blur characterizing image blur based on the difference between the maximum and minimum eigenvalues ​​of the gradient structure tensor matrix includes: for any local window, calculating the ratio of the maximum and minimum eigenvalues ​​of its corresponding gradient structure tensor matrix, and subtracting the difference from 1 to obtain the motion blur; and multiplying the motion blur by the logarithmic operation of the average gray value of the local window as the texture blur of the local window.

[0010] The eigenvalues ​​of the gradient structure tensor matrix reflect the drasticness of gradient changes in a certain direction. In a standard image, if the minimum eigenvalue corresponding to a local window is much smaller than the maximum eigenvalue, it indicates that the problem in the direction corresponding to the minimum eigenvalue is relatively blurry. Furthermore, the smaller the ratio of the minimum to the maximum eigenvalue, the greater the motion blur. Simultaneously calculating the logarithm of the average gray level of the local window allows texture blur to reflect not only gradient direction differences but also the influence of brightness level on motion blur perception.

[0011] Optionally, in the step of selecting local windows with information entropy values ​​greater than a preset entropy threshold as reference blocks, the top 50% of local windows are selected as reference blocks based on the descending order of information entropy corresponding to the local windows.

[0012] A relative threshold-based feature region filtering mechanism was established by selecting the top 50% of local windows as reference patches. This mechanism does not rely on a fixed entropy threshold, but rather performs adaptive filtering based on the information entropy distribution within the same image.

[0013] Optionally, the step of calculating the effective feature density based on the texture blur of the reference patch and the information entropy includes: for any reference patch, the ratio of the square of the information entropy of the reference patch to the texture blur is taken as the effective information content; the ratio between the standard deviation of the gray values ​​in the reference patch and the mean of the gradient magnitude of each pixel in the standard image is taken as the contrast saliency; the product of the effective information content and the contrast saliency is taken as the local feature density; and the mean of the local feature densities of each reference patch is taken as the effective feature density of the standard image.

[0014] The information entropy of the reference patch reflects the amount of information in the reference patch; the larger the value, the greater the effective feature density. Texture blurriness reflects the degree of blurring caused by motion; the larger the value, the fewer features can be extracted from the texture, and the less effective the texture information. Furthermore, by comparing saliency, regions with high information entropy due to noise and pseudo-textures are identified.

[0015] Optionally, the step of calculating the spatial distribution dispersion of the reference map patches based on their distribution characteristics includes: taking the average of the geometric center coordinates of all reference map patches as the high-entropy geometric center; and taking the average of the Euclidean distances from all pixels within the reference map patch to the high-entropy geometric center as the spatial distribution dispersion.

[0016] The high-entropy geometric center reflects the central position of the distribution of all reference tiles. The width of the distribution of reference tiles is reflected by calculating the distance of each pixel to the high-entropy geometric center.

[0017] Optionally, the steps for preprocessing the image to obtain a standard image include: grayscale conversion, median filtering for noise reduction, and contrast-limited adaptive histogram equalization.

[0018] To address the challenges of dusty environments, uneven lighting, and significant color variations in packaging at logistics sites, median filtering effectively removes salt-and-pepper noise while preserving edges. Combined with Limit Contrast Adaptive Histogram Equalization (CLAHE), noise amplification is prevented and the texture details of the packaging surface are highlighted.

[0019] Optionally, the step of adjusting the number of iterations and time step of the AKAZE algorithm by means of an adaptive diffusion factor includes: increasing the number of iterations of the feature extraction algorithm in response to the value of the adaptive diffusion factor being greater than a preset high diffusion threshold; and decreasing the number of iterations of the feature extraction algorithm in response to the value of the adaptive diffusion factor being less than a preset low diffusion threshold.

[0020] For images with an adaptive diffusion factor higher than the high diffusion threshold, the number of iterations is increased to extract more details; for images with an adaptive diffusion factor lower than the low diffusion threshold, the number of iterations is reduced to save computational resources.

[0021] Optionally, the adaptive diffusion factor of multiple standard images can be statistically analyzed, and the 80th percentile of the distribution results can be used as the high diffusion threshold; the 20th percentile of the distribution results can be used as the low diffusion threshold.

[0022] Optionally, the steps of cargo classification include: binarizing the extracted feature points to obtain feature regions; extracting the minimum bounding rectangle of the feature regions, and using the vector formed by the length and width of the minimum bounding rectangle as the feature vector.

[0023] Secondly, this application provides an image classification-based intelligent management system for warehousing and logistics, employing the following technical solution: The image classification-based intelligent warehouse logistics management system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the image classification-based intelligent warehouse logistics management method described above.

[0024] The above-mentioned image classification-based intelligent warehouse logistics management method is generated into a computer program and stored in a memory for loading and execution by a processor. Thus, a system is created based on the memory and processor for easy use.

[0025] This application has the following technical effects: The effective information content of an image is evaluated from multiple dimensions, including texture blur, effective feature density, and spatial distribution dispersion, and an adaptive diffusion iteration factor is synthesized. This factor is used to dynamically adjust the number of iterations and time step of feature extraction, reducing computation when the image is blurry and extracting details when it is clear. This achieves intelligent allocation of computing resources and significantly improves the efficiency of cargo recognition and sorting robustness in complex logistics scenarios. Attached Figure Description

[0026] Figure 1 This is a flowchart of the intelligent warehouse logistics management method based on image classification, as described in this application.

[0027] Figure 2 This is a classification result image of goods identified from a standard image in this application. Detailed Implementation

[0028] This application discloses an intelligent warehouse logistics management method based on image classification. It quantifies the direction and degree of blur by constructing texture blur, and evaluates the amount of truly valuable texture information in an image by constructing an effective feature density. Based on these two dimensions, a diffusion iteration factor that adaptively controls the algorithm's complexity is synthesized. Finally, this factor is used to dynamically adjust the iteration number and time step of the AKAZE (Accelerated KAZE) algorithm. This intelligent scheduling reduces computation when features cannot be extracted due to image blur and increases computation when the image is clear and feature-rich. Ultimately, cargo classification and sorting control are performed based on the extracted features.

[0029] Reference Figure 1 The image classification-based intelligent management method for warehousing and logistics includes steps S1-S5.

[0030] S1: Acquire images of the goods to be sorted on the conveyor equipment and preprocess the images to obtain standard images.

[0031] Implementing this solution first requires building a hardware acquisition system adapted to high-speed logistics environments. Industrial-grade high-speed line scan cameras are installed at key nodes of the logistics conveyor belt, such as directly above the sorting entrance or barcode scanning area. Line scan cameras are chosen because they are suitable for capturing continuously moving long objects, avoiding the inter-frame gaps or overlaps produced by area scan cameras. In conjunction with the line scan cameras, a high-frequency LED strobe light source is placed in the imaging area. The reason for using a strobe light source is that warehouse environments are typically unevenly lit, and goods are in high-speed motion; high-frequency strobe effectively suppresses ambient light interference and minimizes physical motion blur through short exposures. Furthermore, photoelectric trigger sensors are installed on the side of the conveyor belt. When goods pass through the sensing area, the line scan camera is triggered to perform line-by-line scanning, acquiring high-resolution RGB raw images containing the surface texture of the goods, label information, and packaging shape.

[0032] After acquiring the original images, targeted preprocessing is required to adapt them to subsequent algorithm analysis. First, the acquired RGB color images are converted to grayscale images using a weighted average method. This operation aims to reduce the dimensionality of data processing and improve subsequent calculation speed. Simultaneously, considering the typically high dust levels in logistics environments and the potential for conveyor belt vibration to introduce salt-and-pepper noise during image acquisition, a median filtering algorithm is used for image denoising. Compared to mean filtering, median filtering can better preserve edge information while removing isolated noise points, preventing edge smoothing. Subsequently, addressing the issue of significant image contrast differences due to varying packaging colors from different batches of goods, Contrast Limited Adaptive Histogram Equalization (CLAHE) is used to process the grayscale images. CLAHE limits excessive enhancement of local contrast, preventing noise amplification and highlighting the texture details of the packaging surface. Finally, the images are normalized, linearly mapping the pixel values ​​to... The standard range yields a standard image, providing a unified numerical benchmark for subsequent feature index calculations.

[0033] S2: Divide the standard image into multiple local windows; construct a gradient structure tensor matrix for each local window, and construct texture blur based on the difference between the maximum and minimum eigenvalues ​​of the gradient structure tensor matrix to characterize the image blur.

[0034] Even after preprocessing, motion blur due to physical exposure limits may still remain in the standard image. This blur is directional, meaning that textures are stretched and blurred in the direction of motion.

[0035] Specifically, the standard image is uniformly divided into multiple In other embodiments, the standard image can also be divided into overlapping blocks, with the size of the image blocks being the same as the local window, and the overlap length in both the horizontal and vertical directions being the same. For each pixel within a local window, the horizontal gradient is calculated using either the Sobel or Scharr operator. and vertical gradient The Sobel operator is a discrete differential operator used to calculate the approximate gradient of an image's grayscale function.

[0036] Next, the calculated gradients are used to construct the gradient structure tensor matrix for each local window. The structure tensor can well describe the main gradient directions within the local neighborhood.

[0037] Perform eigenvalue decomposition on the gradient structure tensor matrix, calculate its two eigenvalues, and denote the larger of the two eigenvalues ​​as . The smaller eigenvalue is denoted as In a physical sense, The gradient intensity, representing the direction of the most drastic texture change, indicates the clearest texture. It represents the gradient intensity in the direction where the texture change is most gradual, and represents the blurred area caused by motion.

[0038] Based on the above eigenvalues, texture blur is constructed. For any local window, the ratio of the largest eigenvalue to the smallest eigenvalue of its corresponding gradient structure tensor matrix is ​​calculated, and the difference between 1 and this ratio is taken as the motion blur. The product of the motion blur and the logarithmic operation of the average gray value of the local window is taken as the texture blur of the local window.

[0039] The formula for calculating texture blur can be expressed as: In the formula, Indicates the texture blur of a standard image; This represents the total number of standard image blocks; Indicates the first The maximum eigenvalue of the gradient structure tensor within a local window; Indicates the first The minimum eigenvalue of the gradient structure tensor within a local window; To prevent the use of tiny constants with a denominator of zero, the value of this constant is set to 0.1 in this embodiment; Indicates the first The average grayscale value of each window; This represents a logarithmic function with the natural constant as its base.

[0040] When an image has severe motion blur, the gradient along the direction of motion... It will decrease, while the gradient perpendicular to the direction of motion will decrease. Maintaining a large size leads to The value decreases. This method modulates indicators using brightness information. In logistics scenarios, brighter areas (such as white waybills) are more likely to produce visually perceptible motion blur than darker areas (such as cardboard boxes) when in motion. Consequently, the greater the calculated texture blur, the more severe the directional degradation caused by motion blur in the image.

[0041] S3: Select a local window with an information entropy value greater than the preset entropy threshold as a reference tile. Calculate the effective feature density based on the texture fuzziness and information entropy of the reference tile. The effective feature density is negatively correlated with texture fuzziness and positively correlated with information entropy.

[0042] After quantifying the blur level of the standard image, it is also necessary to evaluate the amount of remaining effective information in the image. If the image is severely blurred and lacks texture, blindly performing complex feature extraction is ineffective. Therefore, this step constructs an effective feature density to characterize the richness of texture in the image.

[0043] First, the information entropy of the pixel grayscale values ​​within each local window is calculated. Information entropy reflects the texture richness of an image region. The top 50% of local windows in terms of information entropy are selected as reference patches, representing potential areas of dense feature points. Simultaneously, the average gradient magnitude of the entire image is calculated. As a reference parameter.

[0044] Subsequently, for any reference patch, the ratio of the square of the information entropy of the reference patch to the texture blur is taken as the effective information content, the ratio between the standard deviation of the gray values ​​in the reference patch and the mean of the gradient magnitude of each pixel in the standard image is taken as the contrast saliency, the product of the effective information content and the contrast saliency is taken as the local feature density, and the mean of the local feature densities of each reference patch is taken as the effective feature density of the standard image.

[0045] The formula for calculating the effective feature density can be expressed as: In the formula, The effective feature entropy density of the standard image is represented; the set of reference tiles is used as the reference atlas. The number of reference blocks; Indicates the first reference figure in the reference figure set. Information entropy of a reference image block; For reference, the first one in the atlas Texture blur corresponding to each reference tile; Indicates the first The standard deviation of pixel grayscale values ​​within a reference patch; This represents the average gradient magnitude of the standard image; To prevent the denominator from being zero, the default parameter is set to 0.1.

[0046] In the formula, This represents the texture richness of the reference tile. In the formula, it mainly plays a penalty role. If a reference patch has a high entropy value but large anisotropy of motion texture, then its credibility as an effective feature should be reduced, thereby reducing the value of the final effective feature entropy density.

[0047] The significance of local contrast relative to the global gradient was measured, excluding flat areas in the background that had high entropy values ​​due to noise. Ultimately, the higher the effective feature entropy density, the more the image retains high-density, clear, and effective texture in key areas despite motion, making depth calculation worthwhile.

[0048] S4: Calculate the spatial distribution dispersion of the reference patches based on their distribution characteristics; use the ratio of the width to the height of the standard image as the shape correction coefficient, and use the product of the effective feature density, the reciprocal of the sum of the texture blur of each reference patch, the spatial distribution dispersion, and the shape correction coefficient as the adaptive diffusion factor.

[0049] Statistically analyze the spatial distribution of reference map tiles and calculate the distribution dispersion. Distribution dispersion The steps for obtaining the high-entropy geometric center include: calculating the high-entropy geometric center, which in this embodiment is the mean of the coordinates of the geometric centers of all reference tiles. The average Euclidean distance from the pixels within all reference tiles to the high-entropy geometric center is used as the distribution dispersion. The larger the value, the larger the distribution range of high-entropy regions (potential feature points) in the image.

[0050] Based on this, an adaptive diffusion factor is constructed. The calculation formula is as follows: In the formula: The adaptive diffusion factor represents the standard image; The effective feature density is the effective feature entropy density of a standard image; The spatial distribution dispersion of the reference map patch; The width of the image; The height of the image; Indicates the number of reference tiles in the image; To prevent the parameter from having a denominator of zero, a value of 0.1 is used; For the first in the image Texture blur of a reference tile.

[0051] This indicates the amount of valid information in the standard image; This reflects the breadth of distribution of effective information in a standard image. A larger value means that the image contains a large amount of detail with a wide distribution range, requiring deeper scale-space analysis, thus increasing the adaptive diffusion factor. The aspect ratio is introduced because logistics images are usually obtained by scanning with a line scan camera and are elongated. This is to correct the geometric influence of lateral texture continuity on the diffusion process. This represents the sum of motion texture anisotropy across all local windows in a standard image. The more severe the global motion blur, the more damaged the high-frequency texture information becomes, resulting in diminishing returns and increased risk (potentially leading to oversmoothing) from continuing with numerous iterations. Therefore, it forces... The value decreases. In summary, It is a dimensionless index. The larger the value, the better the image quality and the more worthwhile it is to calculate; The smaller the value, the worse the image quality, and the less computation should be performed.

[0052] S5: Adjust the number of iterations and time step of the AKAZE algorithm by using an adaptive diffusion factor, and extract feature points from the standard image to classify goods.

[0053] This step utilizes the calculated adaptive diffusion factor. The original AKAZE algorithm was improved and sorting control was completed.

[0054] First, it's understandable that the core of the AKAZE algorithm lies in constructing a nonlinear scale space. Its computational complexity primarily depends on the number of iterations and the time step of the nonlinear diffusion. The time step refers to the span between each layer when the algorithm constructs image layers with different levels of blur. The system pre-sets a low diffusion threshold. and high diffusion threshold .

[0055] This embodiment features a low diffusion threshold. and high diffusion threshold Statistical methods were used to determine: several images under normal operating conditions were collected during the initial system deployment phase, and their... The value distribution is set with the 20th percentile as the low diffusion threshold and the 80th percentile as the high diffusion threshold.

[0056] Based on adaptive diffusion factor, the number of iterations and time step are dynamically adjusted: In response to If the image is blurry and lacks information, reduce the number of iterations to decrease computation, and appropriately increase the step size to quickly smooth noise. Specifically, the number of iterations can be adjusted to half of the preset base number of iterations, and the adjustment time should be 1.5 times the preset base step size.

[0057] In response to If the image quality is at a normal level, the standard processing procedure should be maintained, which means processing the standard image using a preset number of basic iterations and a preset basic step size.

[0058] In response to When the image is clear and contains a great deal of detail, the number of iterations is increased to fully extract the image details. In this embodiment, the number of iterations is set to 1.5 times the preset base number of iterations.

[0059] The acquired standard image is input into the improved AKAZE algorithm, which outputs feature points extracted from the image of the goods to be sorted within the standard image. The image is then binarized based on these feature points (feature point positions are set to 1, background to 0). Next, morphological erosion is performed on the binarized image to eliminate discrete noise, and connected component analysis is conducted to obtain several feature regions. The minimum bounding rectangle of each feature region is extracted, and the vector formed by the length and width of this rectangle is used as the feature vector.

[0060] Finally, combining Figure 2 The feature vectors of all feature regions are input into the K-means clustering algorithm. The number of clusters is set. Goods are categorized into four types: large, medium-large, medium-small, and small. The specific classification method involves calculating the magnitude of the central vector of each cluster after clustering, and then sorting them in descending order of magnitude to correspond to the four categories. The identification results are transmitted in real-time to a PLC (Programmable Logic Controller) system. The control system drives sorting wheels or robotic arms according to category instructions to automatically divert goods to designated slots, thus achieving closed-loop intelligent logistics management.

[0061] This application also discloses an image classification-based intelligent warehouse logistics management system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the image classification-based intelligent warehouse logistics management method according to this application.

[0062] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0063] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A warehouse logistics intelligent management method based on image classification, characterized in that, Acquire images of the goods to be sorted on the conveyor equipment, and preprocess the images to obtain standard images; The standard image is divided into multiple local windows; a gradient structure tensor matrix is ​​constructed for each local window, and texture blur is constructed based on the difference between the maximum and minimum eigenvalues ​​of the gradient structure tensor matrix to characterize the image blur. Local windows with information entropy values ​​greater than a preset entropy threshold are selected as reference tiles. Effective feature density is calculated based on the texture fuzziness and information entropy of the reference tiles. Effective feature density is negatively correlated with texture fuzziness and positively correlated with information entropy. Based on the distribution characteristics of the reference patches, the spatial distribution dispersion of the reference patches is calculated; the ratio of the width to the height of the standard image is used as the shape correction coefficient, and the product of the effective feature density, the reciprocal of the sum of the texture blur of each reference patch, the spatial distribution dispersion, and the shape correction coefficient is used as the adaptive diffusion factor. The number of iterations and time step of the AKAZE algorithm are adjusted by an adaptive diffusion factor, and feature points in standard images are extracted for cargo classification.

2. The intelligent warehouse logistics management method based on image classification according to claim 1, characterized in that, The steps for constructing texture blur, which characterizes image blur, based on the difference between the maximum and minimum eigenvalues ​​of the gradient structure tensor matrix include: for any local window, calculating the ratio of the maximum to the minimum eigenvalue of its corresponding gradient structure tensor matrix, and subtracting the difference from 1 to obtain the motion blur; and multiplying the motion blur by the logarithmic operation of the average gray value of the local window to obtain the texture blur of the local window.

3. The intelligent warehouse logistics management method based on image classification according to claim 1, characterized in that, In the step of selecting local windows with information entropy values ​​greater than a preset entropy threshold as reference tiles, the top 50% of local windows are selected as reference tiles based on the descending order of their corresponding information entropy values.

4. The intelligent warehouse logistics management method based on image classification according to claim 1, characterized in that, The steps for calculating the effective feature density based on the texture blur and information entropy of the reference patch include: for any reference patch, the ratio of the square of the information entropy of the reference patch to the texture blur is taken as the effective information content; the ratio between the standard deviation of the gray values ​​in the reference patch and the mean of the gradient magnitude of each pixel in the standard image is taken as the contrast saliency; the product of the effective information content and the contrast saliency is taken as the local feature density; and the mean of the local feature densities of each reference patch is taken as the effective feature density of the standard image.

5. The intelligent warehouse logistics management method based on image classification according to claim 1, characterized in that, Based on the distribution characteristics of the reference map, the steps for calculating the spatial distribution dispersion of the reference map include: taking the average of the geometric center coordinates of all reference map tiles as the high-entropy geometric center; and taking the average of the Euclidean distances from all pixels within the reference map to the high-entropy geometric center as the spatial distribution dispersion.

6. The intelligent warehouse logistics management method based on image classification according to claim 1, characterized in that, The steps for preprocessing an image to obtain a standard image include: grayscale conversion, median filtering for noise reduction, and contrast-limited adaptive histogram equalization.

7. The intelligent warehouse logistics management method based on image classification according to claim 1, characterized in that, The steps of adjusting the number of iterations and time step of the AKAZE algorithm by means of an adaptive diffusion factor include: increasing the number of iterations of the feature extraction algorithm in response to the value of the adaptive diffusion factor being greater than a preset high diffusion threshold; and decreasing the number of iterations of the feature extraction algorithm in response to the value of the adaptive diffusion factor being less than a preset low diffusion threshold.

8. The intelligent warehouse logistics management method based on image classification according to claim 7, characterized in that, The adaptive diffusion factor of multiple standard images was statistically analyzed, and the 80th percentile of the distribution results was used as the high diffusion threshold; the 20th percentile of the distribution results was used as the low diffusion threshold.

9. The intelligent warehouse logistics management method based on image classification according to claim 1, characterized in that, The steps for classifying goods include: binarizing the extracted feature points to obtain feature regions; extracting the minimum bounding rectangle of the feature regions, and using the vector formed by the length and width of the minimum bounding rectangle as the feature vector.

10. A warehouse logistics intelligent management system based on image classification, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the image classification-based intelligent warehouse logistics management method according to any one of claims 1-9.

Images