Warehouse intelligent checking system and method based on deep learning

Abstract

The invention provides an intelligent warehouse checking system and method based on deep learning, and the system comprises an image collection module, an image preprocessing module, a target detection module, and a logic operation module. The image preprocessing module is used for improving the contrast and definition of an image through histogram equalization and image defogging operation, and the target detection module is used for carrying out large-scale training and target detection on the image based on an improved YOLOV10 target detection algorithm so as to obtain the type and quantity of goods; and the logic operation module is used for correcting a result obtained by reasoning of the target detection model by using a logic rule. According to the invention, high-precision cargo detection and inventory can be realized under different illumination conditions and complex environments, the efficiency and accuracy of warehouse management are effectively improved, the method is suitable for various warehouse scenes, and the labor cost and the inventory error can be significantly reduced.

Description

Technical Field

[0001] This application relates to the technical field of warehouse goods inventory, and more specifically, to an intelligent warehouse inventory system and method based on deep learning. Background Art

[0002] Goods inventory in traditional warehouse management usually relies on manual inspection. These methods are inefficient and prone to human errors, thus affecting the accuracy of inventory data. Existing object detection algorithms face many challenges in complex warehouse environments, such as light changes, occlusion, and irregular placement of goods, resulting in false detections or missed detections. It is difficult for the existing technology to provide an efficient and accurate solution to address these issues.

[0003] Currently, most inventory systems on the market are based on solutions such as drones and inspection robots, which are costly and have cumbersome operations. With the popularization of intelligentization and the continuous development of the field of deep learning, there is a need for low-cost and high-efficiency solutions to address the existing situation. Summary of the Invention

[0004] In view of this, this application provides an intelligent warehouse inventory system and method based on deep learning to reduce the inventory cost and improve the inventory efficiency.

[0005] To achieve the above object, in the first aspect, this application provides an intelligent warehouse inventory system based on deep learning, including: an image acquisition module, an image preprocessing module, an object detection module, and a logical operation module. The image acquisition module is used to acquire the images of goods in the warehouse. The image preprocessing module is used to improve the contrast and clarity of the images through histogram equalization and image dehazing operations. The object detection module is used to perform large-scale training and object detection on the images based on an improved YOLOV10 object detection algorithm to obtain the types and quantities of goods. The logical operation module is used to correct the results obtained by the object detection model reasoning using logical rules.

[0006] In the second aspect, this application provides an intelligent warehouse inventory method based on deep learning. The method is applied to the intelligent warehouse inventory system described in the first aspect, and the method includes: Step 1, Image Acquisition: Acquire the images of goods in the warehouse through the image acquisition module and send them to the image preprocessing module; Step 2, Preprocessing: The image preprocessing module performs histogram equalization and image dehazing operations on the acquired images to improve the contrast and clarity of the images; Step 3, Object Detection: The object detection module performs large-scale training and object detection on the preprocessed images based on an improved YOLOV10 object detection algorithm to obtain the types and quantities of goods; Step 4, Logical operation processing: The logical operation module performs post-processing on the target detection results, that is, uses logical rules for correction.

[0007] Further, the histogram equalization step is as follows: (1) Divide the cargo image into R, G, and B channel images; (2) Calculate the histogram for each channel; Let represent the probability that the pixel value in the image is , represent the number of pixels with pixel value , be the total number of pixels, then the calculation formula for the histogram is: ; Where: takes values of 0, 1, 2, ……, L−1, and L is the total number of gray levels; (3) Cumulative distribution function (CDF) calculation formula: ; (4) Calculate the equalized gray value: ; The ; (5) Replace the original gray value of each pixel with the calculated to generate an equalized image; (6) Recombine the equalized R, G, and B channels to generate an equalized RGB image.

[0008] Further, the image defogging step is as follows: (1) Calculate the dark channel For each pixel , its dark channel is defined as: Where, is the value of pixel on color channel c (R, G, B); is a local window centered on ; (2) Atmospheric light estimation Select the top 0.1% of the pixels with the brightest values in the dark channel, and then select the pixel with the maximum brightness in the original image among these pixels as the atmospheric light A; (3) Transmission rate estimation The transmission rate Indicates the intensity of light reaching the camera after passing through the haze and can be estimated by the following formula: where ω is a tuning parameter, is a camera parameter; (4)Transmittance refinement To eliminate the noise in the transmittance estimation, the guided filter method is used to refine the transmittance.

[0009] (5)Restore the image Through the atmospheric light A and the transmittance Restore the defogged image ; where, is a small value infinitely close to zero, used to prevent division by zero.

[0010] Furthermore, the improved YOLOV10 object detection algorithm specifically includes: (1)Feature extraction: Pass the image information through convolutional operations, normalization, SiLU activation function, and the SCINet time series model in sequence to extract the feature information of the image; (2)BackBone network feature fusion: Use the SPPF structure at the last layer of BackBone to splice the downsamplings of different levels of the backbone network to obtain a feature map with richer information; (3)PSA: Pass the feature map fused by SPPF into the PSA module to extract and integrate the spatial information of different scales, thereby establishing a long-term dependence relationship between multi-scale channel attentions; (4)Neck feature fusion: Splice the downsampled feature maps of different levels of the backbone network with the upsampled feature maps of different levels of PAN; (5)Head output head: Splice the feature maps fused by different levels of Neck to three output heads of different sizes to obtain objects of different scales.

[0011] Furthermore, the convolutional operation formula is ; is the pixel value in the image I centered on and is the value of the convolutional kernel.

[0012] Furthermore, the normalization calculation steps include: The first step: Calculate the mean of the mini-batch; ; ​X = {x1, x2, …, xm} is the input of the neural network layer, and m is the size of the mini-batch; Step 2: Calculate the variance of the mini-batch; ; Step 3: Normalization: ; where, is the mean of the mini-batch; is the variance of the mini-batch; is a constant to prevent division by zero; and are learnable parameters used to restore the representational ability of the network; Step 4: Scaling and translation To prevent the standardized BN from restricting the representational ability of the model, scaling and offset are performed on the standardized output; ; where, γ is the learnable scaling parameter and β is the learnable offset parameter.

[0013] Furthermore, the formula of the SiLU activation function is: ; where, is the input.

[0014] Furthermore, the correction formula using logical rules is: ； where, the number of channels in each layer is successively and , i represents the i-th target, and n represents the total number of targets.

[0015] The beneficial effects of this application are: Compared with the prior art, the beneficial effects of this application are: 1. Automatic inventory: Replacing the manual inventory method, realizing the automatic inventory of warehouse goods.

[0016] 2. High-precision detection: Improving the accuracy of target detection through the YOLOV10 target detection algorithm trained on a large scale, and performing excellently in complex environments.

[0017] 3. Cost reduction and efficiency improvement: Greatly reducing the labor time cost, improving the efficiency of goods inventory, and ensuring the safety of goods at the same time. Description of the Drawings

[0018] To more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the accompanying drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present application and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0019] Figure 1 It is an architecture diagram of a warehouse intelligent inventory system based on deep learning.

[0020] Figure 2 It is a flowchart of a warehouse intelligent inventory method based on deep learning. Specific embodiments

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments.

[0022] As Figure 1 shown, a warehouse intelligent inventory system based on deep learning includes: an image acquisition module 110, an image preprocessing module 120, an object detection module 130, and a logical operation module 140. The image acquisition module 110 is used to acquire the images of goods in the warehouse. The image preprocessing module 120 is used to improve the contrast and clarity of the images through histogram equalization and image dehazing operations. The object detection module 130 is used to perform large-scale training and object detection on the images based on an improved YOLOV10 object detection algorithm to obtain the types and quantities of goods. The logical operation module 140 is used to correct the results obtained by the inference of the object detection model using logical rules.

[0023] As Figure 2 shown, the present application provides a warehouse intelligent inventory method based on deep learning. The method is applied to the warehouse intelligent inventory system as Figure 1 shown, and the method includes: Step 1, Image acquisition: Acquire the images of goods in the warehouse through the image acquisition module and send them to the image preprocessing module. The image acquisition module can acquire the images of goods in the warehouse in real time through a camera installed in the warehouse. Step 2, Preprocessing: The image preprocessing module performs histogram equalization and image dehazing operations on the acquired images to improve the contrast and clarity of the images. The image preprocessing module preprocesses the collected images, specifically including histogram equalization processing and image dehazing processing, to improve the contrast and clarity of the images, reduce noise and light interference, thereby enhancing the accuracy of object detection. In practice, opencv can be used to perform histogram equalization and image dehazing operations on the images.

[0024] Step 3, Object Detection: The object detection module performs large-scale training and object detection on the preprocessed images based on the improved YOLOV10 object detection algorithm to obtain the types and quantities of goods. The object detection module detects the preprocessed images based on the improved YOLOV10 algorithm to identify the types and quantities of goods. This algorithm processes data in different dimensions through the SCINet structure and combines the SPPF and PSA modules, enhancing the detection ability in low-light and complex environments.

[0025] Step 4, Logical Operation Processing: The logical operation module post-processes the object detection results, that is, uses logical rules for correction.

[0026] The logical operation module corrects the detection results using custom logical rules, reducing false detections and missed detections, and improving the accuracy of inventory counting.

[0027] Furthermore, the steps of histogram equalization are as follows: (1) Divide the goods image into R, G, and B channel images; Histogram equalization decomposes the image into red, green, and blue channels and performs histogram equalization processing on each channel separately; (2) Calculate the histogram for each channel; Let represent the probability that the pixel value in the image is , represent the number of pixels with pixel value , be the total number of pixels, then the formula for the histogram is: ; where: takes values of 0, 1, 2, ……, L−1, L is the total number of gray levels. For example, for an 8-bit image, L = 256; (3) Cumulative distribution function (CDF) calculation formula: ; (4) Calculate the equalized gray value: ; The is a random function; (5) Replace the original gray value of each pixel with the calculated to generate an equalized image; (6) Recombine the equalized R, G, and B channels to generate an equalized RGB image.

[0028] Further, the image defogging process includes calculating the dark channel, estimating the atmospheric light value, and restoring the image through the transmittance, specifically: (1) Calculate the dark channel For each pixel , its dark channel is defined as: where is the value of pixel on color channel c (R, G, B); is a local window centered on .

[0029] (2) Atmospheric light estimation By calculating the dark channel, the atmospheric light A in the scene can be estimated; usually, the top 0.1% of the pixels with the brightest values in the dark channel are selected, and then the pixel with the largest brightness in the original image among these pixels is selected as the atmospheric light A.

[0030] (3) Transmittance estimation The transmittance represents the intensity of light reaching the camera after passing through the haze and can be estimated by the following formula: where ω is a tuning parameter, usually taking a value of 0.95 to retain a small amount of haze; is the camera parameter; (4) Transmittance refinement To eliminate the noise in the transmittance estimation, methods such as guided filtering are used to refine the transmittance.

[0031] (5) Restore the image Restore the defogged image through the atmospheric light A and the transmittance .

[0032] where is a small value infinitely close to zero, used to prevent division by zero.

[0033] ​Due to the special environment of the warehouse, the quality of the images of goods captured by the camera is low as a whole due to the light and other environmental factors in the warehouse, which is not conducive to the inference of the target detection model. By using histogram equalization and image dehazing operations, the overall quality of the images is improved, and at the same time, it is beneficial to improve the accuracy of the target detection model. Therefore, this application can handle complex warehouse environments such as light changes and occlusions, and maintain high detection performance.

[0034] Furthermore, the improved YOLOV10 target detection algorithm specifically includes: (1) Feature extraction: The image information is processed through convolution operations, normalization, SiLU activation function, and SCINet time series model in sequence to extract the feature information of the image; (2) BackBone network feature fusion: At the last layer of BackBone, the SPPF structure is used to splice the downsamplings of different levels of the backbone network to obtain a feature map with richer information; (3) PSA (Pyramid Slicing Attention): The feature map fused by SPPF is input into the PSA module to extract and integrate spatial information of different scales, so as to establish long-term dependence relationships between multi-scale channel attentions; (4) Neck feature fusion: The downsampling feature maps of different levels of the backbone network are spliced with the upsampling feature maps of different levels of PAN; (5) Head output head: The feature maps fused by different levels of Neck are spliced to three output heads of different sizes to obtain targets of different scales.

[0035] Furthermore, the convolution operation formula is ; is the pixel value centered on in the image I, is the value of the convolution kernel.

[0036] Furthermore, the calculation steps of the normalization (Batch Normalization, abbreviated as BN) include: The first step: Calculate the mean of the mini-batch; ; X = {x1, x2,..., xm} is the input of the neural network layer, and m is the size of the mini-batch; The second step: Calculate the variance of the mini-batch; ; The third step: Normalization (standardization); ; Among them, is the mean of the mini-batch; is the variance of the mini-batch; is a constant to prevent division by zero; and are learnable parameters used to restore the representational ability of the network.

[0037] Step 4: Scaling and translation To prevent the normalization BN from restricting the representational ability of the model, the scaled and offset the output after normalization; ; where γ is the learnable scaling parameter and β is the learnable translation parameter.

[0038] Furthermore, the formula of the SiLU activation function is: ; where is the input.

[0039] This application uses SCINet (a time series model) to improve the YOLOV10 network structure. By processing data in different dimensions through the SCINet structure, different levels of features are obtained, and combined with the SPPF and PSA modules, the detection ability in darker and complex environments is enhanced.

[0040] Furthermore, the correction formula using logical rules is: ； where the number of channels in each layer is successively and ， i represents the i-th target and n represents the total number of targets.

[0041] By customizing the post-processing logical operations, false detections and missed detections are effectively reduced, ensuring the accuracy of the inventory results.

[0042] This application can achieve high-precision cargo detection and inventory under different lighting conditions and complex environments, effectively improving the efficiency and accuracy of warehouse management, applicable to various warehouse scenarios, and can significantly reduce labor costs and inventory errors.

[0043] The above is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed by this application can easily think of changes or substitutions, which should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

1. A warehouse intelligent inventory system based on deep learning, characterized in that: include: An image acquisition module, an image preprocessing module, a target detection module and a logic operation module, wherein the image acquisition module is used to acquire images of goods in the warehouse, the image preprocessing module is used to improve the contrast and clarity of the image through histogram equalization and image defogging operations, the target detection module is used to perform large-scale training and target detection on the image based on the improved YOLOV10 target detection algorithm to obtain the type and quantity of the goods; and the logic operation module is used to correct the results obtained by reasoning the target detection model using logic rules.

2. A warehouse intelligent inventory method based on deep learning, characterized in that: The method is applied to the warehouse intelligent inventory counting system according to claim 1, and the method comprises: Step 1, image acquisition: the image of goods in the warehouse is acquired by the image acquisition module and sent to the image preprocessing module; Step 2, preprocessing: The image preprocessing module performs histogram equalization and image defogging operations on the collected images to improve the contrast and clarity of the images; Step 3, target detection: The target detection module performs large-scale training and target detection on the preprocessed images based on the improved YOLOV10 target detection algorithm to obtain the type and quantity of goods; Step 4: Logical operation processing: The logical operation module performs post-processing on the target detection results, that is, correcting them using logical rules.

3. A warehouse intelligent inventory method based on deep learning as claimed in claim 2, characterized in that: The histogram equalization steps are: (1) Divide the cargo image into three-channel images: R, G, and B; (2) Calculate the histogram for each channel; make Indicates that the pixel value in the image is The probability of Indicates that the pixel value is The number of pixels, is the total number of pixels, the histogram calculation formula is: ； in: The values ​​are 0, 1, 2, …, L−1, where L is the total number of gray levels; (3) Cumulative distribution function (CDF) calculation formula: ； (4) Calculate the grayscale value after equalization: ； Said ; (5) The original gray value of each pixel Replace with the calculated , generate the equalized image; (6) Recombine the equalized R, G, and B channels to generate an equalized RGB image.

4. A warehouse intelligent inventory method based on deep learning as claimed in claim 3, characterized in that: The image defogging steps are: (1) Calculation of dark channel For each pixel , its dark channel Defined as: ； in, It's pixels The value on the color channel c (R, G, B); So Get a local window as the center; (2) Atmospheric light estimation Select the first 0.1% of pixels with the brightest value in the dark channel, and then select the pixel with the largest brightness in the original image among these pixels as the atmospheric light A; (3) Transmittance estimation Transmittance It represents the intensity of light reaching the camera after passing through the haze, which can be estimated by the following formula: ； Among them, ω is a tuning parameter, is the camera parameter; (4) Transmittance refining To remove the noise in the transmittance estimation, the transmittance is refined using a guided filtering method; (5) Image restoration Light A through the atmosphere and transmittance Restore the dehazed image ; ； in, is a small value as close to zero as possible to prevent division by zero.

5. A warehouse intelligent inventory method based on deep learning as claimed in claim 2, characterized in that: The improved YOLOV10 target detection algorithm specifically includes: (1) Feature extraction: The image information is processed through convolution operation, normalization, SiLU activation function and SCINet time series model in sequence to extract the feature information of the image; (2) BackBone network feature fusion: In the last layer of BackBone, the SPPF structure is used to splice the downsampling of different layers of the backbone network to obtain a feature map with richer information; (3) PSA: The feature map fused by SPPF is passed into the PSA module to extract and integrate spatial information of different scales, thereby establishing a long-term dependency relationship between multi-scale channel attentions; (4) Neck feature fusion: concatenate the downsampled feature maps of different layers of the backbone network with the feature maps of different layers of the PAN upsampled feature maps; (5) Head output head: The feature maps fused at different levels of Neck are spliced ​​into three output heads of different sizes to obtain targets of different scales.

6. A warehouse intelligent inventory method based on deep learning as claimed in claim 5, characterized in that: The convolution operation formula is: ; is the image I with The pixel value at the center, is the value of the convolution kernel.

7. A warehouse intelligent inventory method based on deep learning as claimed in claim 6, characterized in that: The standardized calculation step comprises: Step 1: Calculate the mean of mini-batch; ； X={x1,x2,…,xm} is the input of the neural network layer, and m is the size of the mini-batch; Step 2: Calculate the variance of mini-batch; ； Step 3: Normalization: ； in, is the mean of the mini-batch; is the variance of the mini-batch; is a constant to prevent division by zero; and is a learnable parameter used to restore the representation ability of the network; Step 4: Zoom and Pan In order to ensure that the standardized BN does not limit the representation capacity of the model, the standardized output is scaled and offset; ； Among them, γ is a learnable scaling parameter and β is a learnable translation parameter.

8. A warehouse intelligent inventory method based on deep learning as claimed in claim 7, characterized in that: The SiLU activation function formula is: ； in, is the input.

9. A warehouse intelligent inventory method based on deep learning as claimed in claim 8, characterized in that: The correction formula using logical rules for correction is: ； Among them, the number of channels in each layer is and , i represents the i-th target, and n represents the total number of targets.

Images