A warehouse spare parts access method and system based on robot identification

Abstract

This invention discloses a method and system for storing and retrieving warehouse spare parts based on robot recognition, relating to the technical field of intelligent warehousing. The method involves acquiring and executing a preset sampling scheme to obtain an original image set, preprocessing the original image set to obtain a target image set, locating overexposed areas in the target normal image set to obtain positioning tags, fusing the target underexposed image with the target normal image based on the positioning tags to obtain a coarse fused image, post-processing the coarse fused image to obtain a final optimized image, and then recognizing the image to obtain a recognition result. The recognition result is uploaded to a terminal, which determines control commands based on the recognition result and sends the control commands to the robot to perform the warehouse spare parts storage and retrieval operation. By fusing double-exposure images of the same area, overexposed areas in the normal image are accurately replaced, eliminating strong light interference, thereby significantly improving the accuracy and robustness of tag recognition and ensuring the reliability of robot storage and retrieval operations.

Description

Technical Field

[0001] This invention relates to the field of intelligent warehousing technology, specifically to a method and system for storing and retrieving warehouse spare parts based on robot recognition. Background Technology

[0002] With the rapid development of industrial automation technology, automated storage and retrieval of warehouse spare parts has become a core element in improving the efficiency and reliability of warehouse management. Its development has progressed from relying on manual recording and searching in the early stages to adopting robot vision recognition systems based on fixed barcodes or QR codes.

[0003] Publication No.: CN119670783A discloses a cargo label recognition method based on intelligent logistics warehouse management; S1: Using the vertices of a three-dimensional grid as standard points, and any standard point as a collection point, images of the warehouse are collected to determine the target shape in the image; S2: Collect the image to be detected, and determine the shape to be determined in the image to be detected; determine the standard point corresponding to the image to be detected, determine the shape to be compared, and determine the comparison shape based on the Euclidean distance between the center of the comparison shape and the center of the shape to be determined; S3: Determine the similarity between the comparison shape and the shape to be determined, and judge whether the cargo label is an abnormal label based on the similarity.

[0004] In the existing robotic inventory recognition process, the high speed of image acquisition makes strong light and reflections prone to interfering with the images, which reduces the efficiency of subsequent label recognition. Summary of the Invention

[0005] The purpose of this invention is to solve the problem mentioned in the background art that the fast image acquisition speed leads to strong light and reflections easily interfering with the image, which reduces the efficiency of subsequent label recognition. Therefore, this invention proposes a warehouse spare parts storage and retrieval method and system based on robot recognition.

[0006] A first aspect of this invention provides a method for storing and retrieving warehouse spare parts based on robot recognition, the method comprising: A preset sampling scheme is obtained, and the preset sampling scheme is executed to obtain an original image set. The original image set is then preprocessed to obtain a target image set. The target image set consists of multiple image groups. Each image group consists of normal images and low-exposure images of the same region. Overexposed areas are located in the normal target image within the target image group to obtain location tags. Based on the location tags, the underexposed target image is fused with the normal target image to obtain a coarse fused image. The target image group is any one of the target image sets. The coarsely fused image is post-processed to obtain the final optimized image, which is then input into the target model for recognition to obtain the recognition result. The identification results are uploaded to the terminal so that the terminal can determine the control command based on the identification results and send the control command to the robot to perform the warehouse spare parts storage and retrieval operation.

[0007] By employing a multi-exposure fusion repair and adaptive preprocessing mechanism, the challenge of label recognition caused by multiple interferences such as strong light, reflection, dust, and shadows was solved. High-quality reconstruction of label image details under complex lighting conditions was achieved, improving the robustness and accuracy of the recognition model. Furthermore, through a closed-loop process from image acquisition and processing to recognition and decision-making, the automation and high reliability of robot access command generation and execution were ensured, thereby enhancing the operational efficiency and stability of the entire warehouse management system.

[0008] Optionally, fusing the low-exposure image of the target with the normal image of the target based on the positioning tag to obtain a coarse fused image includes: Align the target low-exposure image with the target normal image, locate the target low-exposure image according to the positioning label to obtain a reference region, and extract features from the reference region to obtain reference region features; the reference region features include: gray-level gradient features, geometric parameter features, and pixel connected component features; The target low-exposure image and the target normal image are divided into multiple local blocks. Local blocks without overexposure in the target normal image and corresponding local blocks in the target low-exposure image are selected and combined to obtain training samples. A linear regression equation is constructed based on the training samples. The reference region features are input into a linear regression equation to obtain target features. Based on the target features, overexposed areas in the target normal image are filled with pixels to obtain a coarse fused image.

[0009] By aligning double-exposure images and extracting local geometric features from the low-exposure image, a linear mapping relationship between exposure parameters and label structure is established. This enables high-precision, structure-preserving pixel-level reconstruction of overexposed areas, repairing key geometric features and edge information of barcodes and QR codes lost due to strong light. This fundamentally solves the problem of destroyed recognition features in overexposed areas and improves the accuracy and generalization ability of subsequent model recognition.

[0010] Optionally, post-processing the coarsely fused image to obtain the final optimized image includes: A comprehensive gradient map is obtained by calculating the target low-exposure image using the Sobel operator, the comprehensive gradient map is normalized to obtain a normalized gradient map, and the normalized gradient map is Gaussian filtered to obtain a standard gradient map. A sliding window is obtained, and a window set is obtained by traversing and combining each pixel in the coarse fused image according to the sliding window. A cost function is constructed based on the standard gradient map and the target window. The step size of the sliding window is one pixel. The output value of the target pixel is calculated based on the cost function, and the corresponding pixel in the coarse fused image is replaced according to the target pixel to obtain the final optimized image.

[0011] By constructing an adaptive cost function based on gradient information, pixel-by-pixel optimized filtering is applied to the coarsely fused image. This suppresses artifacts in the filled region and smooths noise while accurately preserving the key edge structure of the label. The geometric restoration results of the previous steps are transformed into images with higher visual quality and more consistent features. This effectively eliminates local inconsistencies and artificial traces that may be introduced by fusion, thus providing clearer and more realistic input for the subsequent recognition model and further improving the overall stability and accuracy of the recognition system.

[0012] Optionally, the final optimized image is input into the target model for recognition to obtain the recognition result. The target model, based on improvements to the YOLOv12n model, includes: The backbone network in the YOLOv12n model is replaced with an improved backbone network, and the output of the improved backbone network is used as the input of the auxiliary detection head to obtain the target model; the YOLOv12n model includes a backbone network and a neck structure; The workflow for improving the backbone network includes: The original image is acquired, and then sequentially input into the Conv module, the Conv module, the Conv module, and the C3k2 module to obtain the first image. The first image is then input into the C3k2 module to obtain the second image. The second image is then sequentially input into the Conv module and the A2C2f module to obtain the third image. Finally, the third image is input into the target Conv module, the A2C2f module, and the feature weighting module to obtain the fourth image. The second, third, and fourth images are used as inputs to the neck structure, and the second, third, and fourth images are used as inputs to the auxiliary detection head.

[0013] By improving the backbone network and introducing an auxiliary detection head, this target model achieves efficient extraction and adaptive fusion of multi-scale features, enhancing its ability to perceive multi-sized labels in the warehouse. The model utilizes the target Conv module to capture local details and global context in parallel, and dynamically focuses on key regions through a feature weighting module, effectively suppressing interference from complex backgrounds. Simultaneously, the multi-level supervision of the auxiliary detection head strengthens the feature representation of small targets and solves the gradient decay problem. While maintaining lightweight computation, it improves the accuracy, robustness, and real-time performance of label recognition under complex lighting conditions, providing a reliable guarantee for precise robot storage and retrieval.

[0014] Optionally, the working principle of the target Conv module includes: Obtain the original tensor, segment the original tensor to obtain the target sub-tensor, and input the target sub-tensor into the 1×1DSConv module, 3×3DSConv module, 5×5DSConv module, 7×7DSConv module and 9×9DSConv module respectively to obtain the target first sub-tensor, target second sub-tensor, target third sub-tensor, target fourth sub-tensor and target fifth sub-tensor; The target first sub-tensor, the target second sub-tensor, the target third sub-tensor, the target fourth sub-tensor, and the target fifth sub-tensor are concatenated along the channel dimension to obtain the first tensor. The first tensor is input into the Conv module to obtain the second tensor. The second tensor is then subjected to batch normalization and activation operations in sequence to obtain the third tensor. The third tensor is then used as the output of the target Conv module.

[0015] Through the collaborative design of the target Conv module and the feature weighting module, an organic combination of multi-scale feature extraction and adaptive feature enhancement is achieved. The target Conv module efficiently captures visual patterns from local details to global context through parallel multi-scale deep separable convolutions, providing a rich feature foundation for label recognition.

[0016] A second aspect of this invention provides a warehouse spare parts storage and retrieval system based on robot recognition, the system comprising: The image acquisition module is used to acquire a preset sampling scheme, execute the preset sampling scheme to obtain an original image set, and preprocess the original image set to obtain a target image set; the target image set consists of multiple image groups; the image group consists of normal images and low-exposure images of the same area; An image fusion module is used to locate overexposed areas in the target normal image in the target image group to obtain location tags, and to fuse the target underexposed image with the target normal image according to the location tags to obtain a coarse fused image; the target image group is any one of the target image sets; The image recognition module is used to post-process the coarsely fused image to obtain the final optimized image, and input the final optimized image into the target model for recognition to obtain the recognition result; The instruction sending module is used to upload the recognition result to the terminal, so that the terminal can determine the control instruction based on the recognition result and send the control instruction to the robot to perform the warehouse spare parts storage and retrieval operation.

[0017] Optionally, the image fusion module includes: An image alignment module is used to align the target low-exposure image with the target normal image, locate the target low-exposure image according to the positioning label to obtain a reference region, and extract features from the reference region to obtain reference region features; the reference region features include: gray-level gradient features, geometric parameter features, and pixel connected component features; The image segmentation module is used to segment the target low-exposure image and the target normal image to obtain multiple local blocks, select local blocks without overexposure in the target normal image and the corresponding local blocks in the target low-exposure image to combine to obtain training samples, and construct a linear regression equation based on the training samples; The image filling module is used to input the reference region features into a linear regression equation to obtain target features, and to fill the overexposed areas in the target normal image with pixels according to the target features to obtain a coarse fused image.

[0018] Optionally, the image recognition module includes: The image processing module is used to calculate a comprehensive gradient map on the target low-exposure image using the Sobel operator, normalize the comprehensive gradient map to obtain a normalized gradient map, and perform Gaussian filtering on the normalized gradient map to obtain a standard gradient map. The pixel traversal module is used to obtain a sliding window, traverse and combine each pixel in the coarse fused image according to the sliding window to obtain a window set, and construct a cost function based on the standard gradient map and the target window; the step size of the sliding window is one pixel. The pixel replacement module is used to calculate the output value of the target pixel according to the cost function, and replace the corresponding pixel in the coarse fused image according to the target pixel to obtain the final optimized image.

[0019] Optionally, the final optimized image is input into the target model for recognition to obtain the recognition result. The target model, based on improvements to the YOLOv12n model, includes: The backbone network in the YOLOv12n model is replaced with an improved backbone network, and the output of the improved backbone network is used as the input of the auxiliary detection head to obtain the target model; the YOLOv12n model includes a backbone network and a neck structure; The workflow for improving the backbone network includes: The original image is acquired, and then sequentially input into the Conv module, the Conv module, the Conv module, and the C3k2 module to obtain the first image. The first image is then input into the C3k2 module to obtain the second image. The second image is then sequentially input into the Conv module and the A2C2f module to obtain the third image. Finally, the third image is input into the target Conv module, the A2C2f module, and the feature weighting module to obtain the fourth image. The second, third, and fourth images are used as inputs to the neck structure, and the second, third, and fourth images are used as inputs to the auxiliary detection head.

[0020] Optionally, the working principle of the target Conv module includes: Obtain the original tensor, segment the original tensor to obtain the target sub-tensor, and input the target sub-tensor into the 1×1DSConv module, 3×3DSConv module, 5×5DSConv module, 7×7DSConv module and 9×9DSConv module respectively to obtain the target first sub-tensor, target second sub-tensor, target third sub-tensor, target fourth sub-tensor and target fifth sub-tensor; The target first sub-tensor, the target second sub-tensor, the target third sub-tensor, the target fourth sub-tensor, and the target fifth sub-tensor are concatenated along the channel dimension to obtain the first tensor. The first tensor is input into the Conv module to obtain the second tensor. The second tensor is then subjected to batch normalization and activation operations in sequence to obtain the third tensor. The third tensor is then used as the output of the target Conv module.

[0021] The beneficial effects of this invention are: This invention proposes a method for storing and retrieving warehouse spare parts based on robot recognition. By acquiring normal and low-exposure images of the same area, and utilizing the details preserved in the low-exposure image that are not affected by overexposure, the overexposed areas in the normal image caused by strong light or reflection are accurately located, replaced, and fused to generate a coarsely fused image with optimized quality. This image is then post-processed and input into the model for recognition. This method improves the accuracy and robustness of label recognition in complex lighting warehouse environments, ensuring the reliability and efficiency of subsequent robot storage and retrieval operations. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a method for storing and retrieving warehouse spare parts based on robot recognition, provided as an embodiment of the present invention; Figure 2 A model structure diagram of a warehouse spare parts storage and retrieval method based on robot recognition provided in an embodiment of the present invention; Figure 3 This is a module structure diagram of a warehouse spare parts storage and retrieval method based on robot recognition provided in an embodiment of the present invention. Detailed Implementation

[0023] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0024] This invention provides a method for storing and retrieving warehouse spare parts based on robot recognition. See also... Figure 1 , Figure 1 A flowchart illustrating a method for storing and retrieving warehouse spare parts based on robot recognition, provided in an embodiment of the present invention. The method includes the following steps: S101, Obtain a preset sampling scheme, execute the preset sampling scheme to obtain the original image set, and preprocess the original image set to obtain the target image set; S102, locate the overexposed area of ​​the target normal image in the target image group to obtain the location tag, and fuse the target underexposed image with the target normal image according to the location tag to obtain a coarse fused image. S103, Post-process the coarsely fused image to obtain the final optimized image, and input the final optimized image into the target model for recognition to obtain the recognition result; S104, the recognition result is uploaded to the terminal so that the terminal can determine the control command based on the recognition result and send the control command to the robot to perform the warehouse spare parts storage and retrieval operation.

[0025] The target image set consists of multiple image groups; each image group consists of normal images and low-exposure images of the same region; the target image group is any one of the images in the target image set. The present invention provides a method for storing and retrieving warehouse spare parts based on robot recognition. By fusing double-exposure images of the same area, the overexposed areas in the normal image are accurately replaced, and strong light interference is eliminated, thereby significantly improving the accuracy and robustness of label recognition and ensuring the reliability of robot storage and retrieval operations.

[0026] In one implementation, a preset sampling scheme is obtained. The preset sampling scheme is as follows: the robot first moves to the spare parts label area in the warehouse, at a distance of 30~50cm from the label, and within a shooting angle of ±15°, triggering the camera to continuously capture 2 frames of images, finally obtaining the original image set; normal exposure images: the baseline exposure time is preset according to the label material, for example: 50ms for paper, 40ms for plastic, and 30ms for metal, capturing the complete geometric features of non-overexposed areas, such as barcode bars and spaces, QR code positioning angles, etc.; low exposure images: the initial exposure time is set to 1 / 5~1 / 8 of the normal exposure time, and the saturated pixel ratio is checked after acquisition. If it is >1%, it is fine-tuned to ≤1% to ensure that the unsaturated grayscale gradient information is retained in the overexposed areas.

[0027] In one implementation, the original image set is preprocessed to obtain the target image set. The two sets of acquired images are then processed separately to eliminate composite interference from the warehouse environment, providing high-quality data for subsequent restoration: Dust and oil stain suppression: Adaptive median filtering is used, dynamically adjusting the window size (3×3~7×7). When a pixel grayscale change exceeds 50, the window is enlarged to filter noise; when the edge gradient is ≥30, the window remains 3×3 to avoid edge blurring. Shadow removal: The Retinex algorithm is used to decompose the image's reflection and illumination components, correcting shadow areas with an illumination component ratio ≤0.3, ensuring the overall label brightness variance is ≤20, avoiding confusion between shadows and overexposure boundaries. Color crosstalk elimination, for color labels: Under low ambient light (<50 lux), the RGB channel coupling matrix A is calculated by projecting pure red, green, and blue images, and the inverse matrix A is used... -1 Correct pixel values ​​to control inter-channel interference error to ≤3%.

[0028] In one implementation, the warehouse terminal is, for example, a warehouse management system (WMS) server, a scheduling console, or an industrial control computer. After receiving the identification result, the terminal compares it with the inventory database and automatically generates or assists in generating specific storage and retrieval instructions, such as material location, quantity, and target shelf. The terminal then sends the instructions to the corresponding robot, such as an AGV, via an industrial wireless network within the warehouse, such as Wi-Fi 6 or a 5G private network, driving it to perform navigation, robotic arm grasping, or shelf handling operations, thereby completing the automated storage and retrieval task.

[0029] In one embodiment, fusing the low-exposure image of the target with the normal image of the target based on the location tag to obtain a coarse fused image includes: Align the low-exposure target image with the normal target image, locate the low-exposure target image based on the positioning label to obtain the reference region, and extract features from the reference region to obtain the reference region features; the reference region features include: gray-level gradient features, geometric parameter features, and pixel connected component features; The target low-exposure image and the target normal image are divided into multiple local blocks. The local blocks without overexposure in the target normal image and the corresponding local blocks in the target low-exposure image are selected to form training samples. A linear regression equation is constructed based on the training samples. The reference region features are input into the linear regression equation to obtain the target features. Based on the target features, the overexposed areas in the normal image of the target are filled with pixels to obtain a coarse fused image.

[0030] In one implementation, the target low-exposure image is aligned with the target normal image. For the two pre-processed images, normal exposure and low exposure, feature points are extracted using the SIFT algorithm, such as the corners of the label, the edges of the barcode bars, and the contours of the QR code positioning angles. At least 50 feature points are extracted for each image to ensure matching stability. Feature point matching: The similarity of feature points is calculated using Euclidean distance. Matching pairs with a distance less than a threshold (set to 1.5 times the minimum distance) are retained, and mismatched points are removed, such as the features of debris in the warehouse background. Homography matrix solution: Based on the successfully matched feature points, the homography matrix H (3×3 matrix) is fitted using the RANSAC algorithm. The number of iterations is set to 1000, and the inlier threshold is set to 2 pixels to ensure that the matrix can accurately describe the spatial mapping relationship between the two images. Image transformation: The low-exposure image is transformed through the homography matrix H to make the label areas of the two images completely aligned, with a final displacement deviation ≤ 1 pixel, adapting to the slight shaking during robot shooting.

[0031] In one implementation, a reference region is obtained by locating the target low-exposure image based on the positioning tag. Since the normal image and the low-exposure image are already aligned, the corresponding position in the target low-exposure image can be found based on the positioning tag. Feature extraction is performed on the reference region to obtain its features, including: grayscale gradient features: the Sobel operator (x, y directions) is used to calculate the edge gradient of the reference region, with a threshold of 20. Pixels with a gradient ≥ 20 are retained, i.e., the bar / space edges of the tag and the contour edges of the positioning corners, and their coordinates and gradient directions are recorded. Geometric parameter features: Barcode: Based on the alternation of bars and spaces, the average width of the bars and the average width of the spaces within the reference region are statistically analyzed, and the bar / space width ratio (e.g., bar width:space width = 1:2) is calculated. The contour shapes of the start and end symbols are also recorded. QR code: The rectangular contour of the positioning corner within the reference region is detected, and the side length, interior angle (ideally 90°), center coordinates, and auxiliary contour features around the positioning corner are calculated.

[0032] In one implementation, the target underexposed image and the target normal image are divided into multiple local blocks. The label area of ​​the normal exposed image (target normal image) is evenly divided into 3×3 local blocks, each block having a size of ≥10×10 pixels. If the label size is small, it is adjusted to 2×2 blocks. Only the local blocks containing the overexposed area are processed, and non-overexposed blocks are skipped directly, saving computing power. Training samples are obtained by combining unexposed local patches from the target normal exposure image and their corresponding local patches from the target low exposure image. A linear regression equation is constructed based on these training samples: Unexposed local patches from the normal exposure image and their corresponding patches in the low exposure image are selected as training samples (each sample patch contains ≥30 pixels of geometric feature data); a geometric feature regression equation is established: using features of the low exposure image as the independent variable X (e.g., bar-to-space ratio, gradient value) and features of the normal exposure image as the dependent variable Y, a linear regression model is constructed: Y = αX + β; coefficients are solved using the least squares method to solve for the coefficients α (slope) and β (intercept), and the goodness of fit R is required. 2 ≥0.95 (refer to the linear correlation standard in the paper) to ensure that the linear regression equation can map the feature relationship between the two sets of images.

[0033] In one implementation, the reference region features are input into a linear regression equation to obtain target features. Based on the target features, pixel filling is performed on the overexposed areas in the normal target image to obtain a coarse fused image. Pixel filling is as follows: Barcode: Based on the target bar-to-space width ratio, pixel grayscale is filled in the overexposed areas, with bar grayscale ≤ 60 and space grayscale ≥ 190, ensuring that the bar-to-space edges are aligned with the surrounding non-overexposed areas (edge ​​deviation ≤ 1 pixel); QR code: Based on the target positioning angle size and center coordinates, the rectangular outline of the positioning angle is filled (grayscale ≤ 50), while restoring the auxiliary pattern around the positioning angle, ensuring that the aspect ratio error of the positioning angle is ≤ 3%; Character label: Based on the target character outline gradient, character strokes are filled (grayscale ≤ 70), ensuring stroke continuity with no breaks exceeding 2 pixels in length; Boundary transition processing: At the boundary (width 2 pixels) between the overexposed and non-overexposed areas, a grayscale gradient is used for filling, such as from 60 grayscale in the overexposed area to 50 grayscale in the non-overexposed area, to avoid obvious boundaries.

[0034] In one embodiment, post-processing the coarsely fused image to obtain the final optimized image includes: The Sobel operator is used to calculate the composite gradient map of the low-exposure image of the target. The composite gradient map is then normalized to obtain the normalized gradient map. Finally, the normalized gradient map is subjected to Gaussian filtering to obtain the standard gradient map. Obtain a sliding window, and use the sliding window to traverse and combine each pixel in the coarsely fused image to obtain a window set. Construct a cost function based on the standard gradient map and the target window. The step size of the sliding window is one pixel. The output value of the target pixel is calculated based on the cost function, and the corresponding pixel in the coarse fused image is replaced according to the target pixel to obtain the final optimized image.

[0035] In one implementation, the Sobel operator is used to calculate the gradients in the x and y directions, using the formula... The comprehensive gradient map is obtained, where G is the comprehensive gradient map, G x G is the gradient map in the x-direction. y The gradient map is shown in the y-direction. The composite gradient map is then normalized: gradient values ​​are normalized to the range of 0-255 to highlight the difference between the label's bar / space edges, positioning angle contours (high gradient values), and the background (low gradient values). A Gaussian filter is then applied to the normalized gradient map to obtain a standard gradient map: a light Gaussian filter (σ=0.8, 3×3 window) is applied to the gradient map to eliminate minor noise while preserving core edge features.

[0036] In one implementation, the sliding window is set to 5×5, the window sliding step size is set to 1 pixel, and optimization is performed pixel by pixel; the cost function is constructed based on the standard gradient map and the target window, and the cost function is: , where ω k For a local sliding window centered at pixel k, the i-window ω k any pixel within, a k For window ω k The linear slope coefficient within, b k For window ω k The linear intercept coefficient within I i p represents the grayscale value of the standard gradient map at pixel i. i Let i be the grayscale value of the coarsely blended image at pixel i. The regularization parameter z is set to 0.01, (a k *I i +b k -p i ) 2 This is the reconstruction error term between the standard gradient map and the coarsely fused image. This is a regularization constraint term.

[0037] In one implementation, the output value of the target pixel is calculated based on a cost function. , where q i The final filtered output value of pixel i, Ω i The set of all windows containing pixel i, where N is the set Ω. i The number of windows; the corresponding pixels in the coarsely fused image are replaced according to the target pixel to obtain the final optimized image.

[0038] In one embodiment, the final optimized image is input into the target model for recognition to obtain the recognition result. The target model is based on improvements to the YOLOv12n model, including: The backbone network in the YOLOv12n model is replaced with an improved backbone network, and the output of the improved backbone network is used as the input of the auxiliary detection head to obtain the target model; the YOLOv12n model includes a backbone network and a neck structure; Improving the backbone network workflow includes: The original image is obtained and sequentially input into the Conv module, Conv module, Conv module and C3k2 module to obtain the first image. The first image is input into the C3k2 module to obtain the second image. The second image is sequentially input into the Conv module and A2C2f module to obtain the third image. The third image is input into the target Conv module, A2C2f module and feature weighting module to obtain the fourth image. Images 2, 3, and 4 are used as inputs to the neck structure, and images 2, 3, and 4 are used as inputs to the auxiliary detection head.

[0039] In one implementation, see [link to implementation details]. Figure 2 , Figure 2 This is a model structure diagram of a warehouse spare parts storage and retrieval method based on robot recognition provided in an embodiment of the present invention. By replacing the backbone network with an improved backbone network integrating a target Conv module, the model achieves parallel extraction and adaptive fusion of multi-scale features. The 1×1 to 9×9 depthwise separable convolutional layers in the target Conv module capture local details and global contextual information respectively, while the dynamic weighting mechanism enables the model to automatically adjust the contribution ratio of features at each scale according to the target size, enhancing the model's ability to perceive multi-scale targets in warehouse label recognition. At the same time, due to the use of lightweight depthwise separable convolution and parameter optimization, the detection accuracy is improved while maintaining low computational overhead, supporting the real-time inference requirements of high frame rate.

[0040] In one implementation, an improved feature weighting module introduced at the end of the backbone network calculates attention weights in three dimensions: height, width, and channel. This achieves synergistic enhancement of spatial and channel information, enabling the model to effectively suppress interference from complex backgrounds and focus on key target regions. Simultaneously, the intermediate feature maps of the backbone network (images 2, 3, and 4) are input to both the neck structure and the auxiliary detection head. The auxiliary supervision signal strengthens the training gradient of low-level features, mitigating the weakening of small target feature representations due to gradient decay.

[0041] In one embodiment, the target Conv module operates as follows: Obtain the original tensor, segment the original tensor to obtain the target sub-tensor, and input the target sub-tensor into the 1×1DSConv module, 3×3DSConv module, 5×5DSConv module, 7×7DSConv module and 9×9DSConv module respectively to obtain the target first sub-tensor, target second sub-tensor, target third sub-tensor, target fourth sub-tensor and target fifth sub-tensor; The first, second, third, fourth, and fifth subtensors of the target are concatenated along the channel dimension to obtain the first tensor. The first tensor is then input into the Conv module to obtain the second tensor. The second tensor is then subjected to batch normalization and activation operations in sequence to obtain the third tensor. The third tensor is then used as the output of the target Conv module. In one implementation, the working principle of the feature weighting module includes: Obtain the original feature tensor, perform height-dimensional pooling on the original feature tensor to obtain the height feature tensor, perform width-dimensional pooling on the original feature tensor to obtain the width feature tensor, and perform channel-dimensional pooling on the original feature tensor to obtain the channel feature tensor. The height weights are obtained by performing convolution and activation operations on the height feature tensor in sequence; the width weights are obtained by performing convolution and activation operations on the width feature tensor in sequence; and the channel weights are obtained by performing convolution and activation operations on the channel feature tensor in sequence. The height weight is multiplied element-wise with the height feature tensor to obtain the height-weighted tensor. The width weight is multiplied element-wise with the width feature tensor to obtain the width-weighted tensor. The channel weight is multiplied element-wise with the channel feature tensor to obtain the channel-weighted tensor. The height-weighted tensor, width-weighted tensor, and channel-weighted tensor are fused to obtain the enhanced feature tensor. The enhanced feature tensor is used as the output of the feature weighting module.

[0042] In one implementation, see [link to implementation details]. Figure 3 , Figure 3This is a module structure diagram of a warehouse spare parts storage and retrieval method based on robot recognition provided in an embodiment of the present invention. Segmenting the original tensor allows different sub-tensors to be processed independently, reducing the computational burden of a single path and laying a structural foundation for subsequent multi-scale extraction. Subsequently, depthwise separable convolution (DSConv) modules of 1×1 to 9×9 are applied to each sub-tensor: small convolution kernels (such as 1×1, 3×3) focus on local details and texture features, suitable for small target perception; large convolution kernels (such as 7×7, 9×9) capture a wide range of context and global structural information, improving the representation ability of large targets. The concatenation operation integrates features of these five scales in the channel dimension, forming a rich multi-scale feature set; the subsequent 1×1 convolution module performs channel compression and information fusion on the concatenated features, reducing redundancy; batch normalization stabilizes the training process and accelerates convergence, while the activation function introduces nonlinearity, enhancing the expressive power of the features.

[0043] In one implementation, parallel processing of segmentation and multi-scale DSConv enables the model to simultaneously capture visual patterns from local to global perspectives, overcoming the limitations of a single receptive field. The stitching operation preserves independent information at each scale, avoiding feature confusion. Subsequent convolution and normalization steps intelligently integrate multi-scale features, outputting a unified feature representation that combines detail sensitivity and context awareness. This improves the model's detection accuracy for multi-scale targets (such as QR codes and text) in warehouse label recognition while maintaining lightweight computational characteristics and supporting real-time inference.

[0044] In one implementation, the synergistic effect of three-dimensional pooling, adaptive weight generation, and full-dimensional weighted fusion enhances the representational and discriminative power of features. Soft pooling operations in height, width, and channel dimensions extract key statistical information on spatial structure and channel response, providing fine-grained feature summaries for attention weight calculation. Dimension-specific weights generated through convolution and activation enable the model to adaptively learn and highlight key regions and channels relevant to the target in each dimension. Element-wise multiplication and fusion of the learned weights with the original features achieves an organic unity of spatial and channel attention, allowing the output features to simultaneously suppress complex background interference and enhance the response of the target region. This enables the model to more accurately focus on key features of multi-scale targets in complex scenarios such as warehouse label recognition, effectively improving the robustness and accuracy of detection.

[0045] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should still fall within the scope of the claims of the present invention.

Claims

1. A method for storing and retrieving warehouse spare parts based on robot recognition, characterized in that, The method includes: A preset sampling scheme is obtained, and the preset sampling scheme is executed to obtain an original image set. The original image set is then preprocessed to obtain a target image set. The target image set consists of multiple image groups. Each image group consists of normal images and low-exposure images of the same region. Overexposed areas are located in the normal target image within the target image group to obtain location tags. Based on the location tags, the underexposed target image is fused with the normal target image to obtain a coarse fused image. The target image group is any one of the target image sets. The coarsely fused image is post-processed to obtain the final optimized image, which is then input into the target model for recognition to obtain the recognition result. The identification results are uploaded to the terminal so that the terminal can determine the control command based on the identification results and send the control command to the robot to perform the warehouse spare parts storage and retrieval operation.

2. The method for storing and retrieving warehouse spare parts based on robot recognition according to claim 1, characterized in that, The coarse fused image is obtained by fusing the low-exposure image of the target with the normal image of the target based on the positioning tag, including: Align the target low-exposure image with the target normal image, locate the target low-exposure image according to the positioning label to obtain a reference region, and extract features from the reference region to obtain reference region features; the reference region features include: gray-level gradient features, geometric parameter features, and pixel connected component features; The target low-exposure image and the target normal image are divided into multiple local blocks. Local blocks without overexposure in the target normal image and corresponding local blocks in the target low-exposure image are selected and combined to obtain training samples. A linear regression equation is constructed based on the training samples. The reference region features are input into a linear regression equation to obtain target features. Based on the target features, overexposed areas in the target normal image are filled with pixels to obtain a coarse fused image.

3. The method for storing and retrieving warehouse spare parts based on robot recognition according to claim 1, characterized in that, Post-processing the coarsely fused image to obtain the final optimized image includes: A comprehensive gradient map is obtained by calculating the target low-exposure image using the Sobel operator, the comprehensive gradient map is normalized to obtain a normalized gradient map, and the normalized gradient map is Gaussian filtered to obtain a standard gradient map. A sliding window is obtained, and a window set is obtained by traversing and combining each pixel in the coarse fused image according to the sliding window. A cost function is constructed based on the standard gradient map and the target window. The step size of the sliding window is one pixel. The output value of the target pixel is calculated based on the cost function, and the corresponding pixel in the coarse fused image is replaced according to the target pixel to obtain the final optimized image.

4. The warehouse spare parts storage and retrieval method based on robot recognition according to claim 1, characterized in that, The final optimized image is input into the target model for recognition to obtain the recognition result. The target model is based on improvements to the YOLOv12n model, including: The backbone network in the YOLOv12n model is replaced with an improved backbone network, and the output of the improved backbone network is used as the input of the auxiliary detection head to obtain the target model; the YOLOv12n model includes a backbone network and a neck structure; The workflow of the improved backbone network includes: The original image is acquired, and then sequentially input into the Conv module, the Conv module, the Conv module, and the C3k2 module to obtain the first image. The first image is then input into the C3k2 module to obtain the second image. The second image is then sequentially input into the Conv module and the A2C2f module to obtain the third image. Finally, the third image is input into the target Conv module, the A2C2f module, and the feature weighting module to obtain the fourth image. The second, third, and fourth images are used as inputs to the neck structure, and the second, third, and fourth images are used as inputs to the auxiliary detection head.

5. The warehouse spare parts storage and retrieval method based on robot recognition according to claim 4, characterized in that, The working principle of the target Conv module includes: Obtain the original tensor, segment the original tensor to obtain the target sub-tensor, and input the target sub-tensor into the 1×1DSConv module, 3×3DSConv module, 5×5DSConv module, 7×7DSConv module and 9×9DSConv module respectively to obtain the target first sub-tensor, target second sub-tensor, target third sub-tensor, target fourth sub-tensor and target fifth sub-tensor; The target first sub-tensor, the target second sub-tensor, the target third sub-tensor, the target fourth sub-tensor, and the target fifth sub-tensor are concatenated along the channel dimension to obtain the first tensor. The first tensor is input into the Conv module to obtain the second tensor. The second tensor is then subjected to batch normalization and activation operations in sequence to obtain the third tensor. The third tensor is then used as the output of the target Conv module.

6. A warehouse spare parts storage and retrieval system based on robot recognition, characterized in that, The system includes: The image acquisition module is used to acquire a preset sampling scheme, execute the preset sampling scheme to obtain an original image set, and preprocess the original image set to obtain a target image set; the target image set consists of multiple image groups; the image group consists of normal images and low-exposure images of the same area; An image fusion module is used to locate overexposed areas in the target normal image in the target image group to obtain location tags, and to fuse the target underexposed image with the target normal image according to the location tags to obtain a coarse fused image; the target image group is any one of the target image sets; The image recognition module is used to post-process the coarsely fused image to obtain the final optimized image, and input the final optimized image into the target model for recognition to obtain the recognition result; The instruction sending module is used to upload the recognition result to the terminal, so that the terminal can determine the control instruction based on the recognition result and send the control instruction to the robot to perform the warehouse spare parts storage and retrieval operation.

7. A warehouse spare parts storage and retrieval system based on robot recognition according to claim 6, characterized in that, The image fusion module includes: An image alignment module is used to align the target low-exposure image with the target normal image, locate the target low-exposure image according to the positioning label to obtain a reference region, and extract features from the reference region to obtain reference region features; the reference region features include: gray-level gradient features, geometric parameter features, and pixel connected component features; The image segmentation module is used to segment the target low-exposure image and the target normal image to obtain multiple local blocks, select local blocks without overexposure in the target normal image and the corresponding local blocks in the target low-exposure image to combine to obtain training samples, and construct a linear regression equation based on the training samples; The image filling module is used to input the reference region features into a linear regression equation to obtain target features, and to fill the overexposed areas in the target normal image with pixels according to the target features to obtain a coarse fused image.

8. A warehouse spare parts storage and retrieval system based on robot recognition according to claim 6, characterized in that, The image recognition module includes: The image processing module is used to calculate a comprehensive gradient map on the target low-exposure image using the Sobel operator, normalize the comprehensive gradient map to obtain a normalized gradient map, and perform Gaussian filtering on the normalized gradient map to obtain a standard gradient map. The pixel traversal module is used to obtain a sliding window, traverse and combine each pixel in the coarse fused image according to the sliding window to obtain a window set, and construct a cost function based on the standard gradient map and the target window; the step size of the sliding window is one pixel. The pixel replacement module is used to calculate the output value of the target pixel according to the cost function, and replace the corresponding pixel in the coarse fused image according to the target pixel to obtain the final optimized image.

9. A warehouse spare parts storage and retrieval system based on robot recognition according to claim 6, characterized in that, The final optimized image is input into the target model for recognition to obtain the recognition result. The target model is based on improvements to the YOLOv12n model, including: The backbone network in the YOLOv12n model is replaced with an improved backbone network, and the output of the improved backbone network is used as the input of the auxiliary detection head to obtain the target model; the YOLOv12n model includes a backbone network and a neck structure; The workflow of the improved backbone network includes: The original image is acquired, and then sequentially input into the Conv module, the Conv module, the Conv module, and the C3k2 module to obtain the first image. The first image is then input into the C3k2 module to obtain the second image. The second image is then sequentially input into the Conv module and the A2C2f module to obtain the third image. Finally, the third image is input into the target Conv module, the A2C2f module, and the feature weighting module to obtain the fourth image. The second, third, and fourth images are used as inputs to the neck structure, and the second, third, and fourth images are used as inputs to the auxiliary detection head.

10. A warehouse spare parts storage and retrieval system based on robot recognition according to claim 9, characterized in that, The working principle of the target Conv module includes: Obtain the original tensor, segment the original tensor to obtain the target sub-tensor, and input the target sub-tensor into the 1×1DSConv module, 3×3DSConv module, 5×5DSConv module, 7×7DSConv module and 9×9DSConv module respectively to obtain the target first sub-tensor, target second sub-tensor, target third sub-tensor, target fourth sub-tensor and target fifth sub-tensor; The target first sub-tensor, the target second sub-tensor, the target third sub-tensor, the target fourth sub-tensor, and the target fifth sub-tensor are concatenated along the channel dimension to obtain the first tensor. The first tensor is input into the Conv module to obtain the second tensor. The second tensor is then subjected to batch normalization and activation operations in sequence to obtain the third tensor. The third tensor is then used as the output of the target Conv module.

Images