Two-dimensional code image signal processing method and system for AGV navigation

Abstract

The invention discloses a two-dimensional code image signal processing method and system for AGV navigation. The method comprises the following steps: acquiring an original image containing a two-dimensional code; noise suppression, contrast reconstruction and local structure compensation are carried out through an image restoration module, and an enhanced image is output; a two-dimensional code positioning module is used for determining two-dimensional code corner points or boundaries, and a standardized two-dimensional code area image is obtained through geometric correction; inputting the standardized image into a decoding module, performing semantic feature extraction and sequence prediction through a deep neural network, and directly outputting coding information and confidence; and finally, the information is transmitted to an AGV control system for navigation decision making. The system correspondingly comprises an image acquisition module, a restoration module, a positioning correction module, a decoding module and an interface module, all the modules are subjected to lightweight optimization to adapt to an embedded platform, and the processing delay is 30 ms. According to the invention, complex industrial scenes such as oil stains, dust and defects can be effectively handled, and the robustness of two-dimensional code recognition and the reliability of AGV navigation are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and in particular relates to a method and system for processing QR code image signals for AGV navigation. Background Technology

[0002] In the fields of intelligent manufacturing and warehousing logistics, Automated Guided Vehicles (AGVs / AMRs) widely use two-dimensional barcodes (such as QR Codes and DataMatrix) laid on the ground for high-precision positioning and command acquisition. The traditional barcode reading process usually includes steps such as image acquisition, image preprocessing, binarization, locator detection, perspective correction, grid sampling, and error correction decoding.

[0003] However, existing technologies have significant shortcomings in several industrial operating conditions: 1. Limited image restoration capabilities based on rules: Traditional filtering, super-resolution, or deblurring methods are effective for mild blurring / low resolution, but when QR codes are covered by oil, dust, or physically torn, resulting in large-area missing modules, it is difficult to complete the missing texture at the semantic level, leading to backend error correction failure; some generative model-based work can improve visual effects, but it does not specifically constrain machine readability, and the generated results may not be conducive to standard decoder reading.

[0004] 2. Geometric localization is strongly dependent on locators: Traditional methods rely on Finder Pattern or Timing Pattern for localization; when locators are worn out or missing, localization based on geometric rules will fail, and the entire decoding process will be interrupted.

[0005] 3. Modular processing leads to information loss: If enhancement, localization and decoding are each independent modules, it is difficult for the backend to compensate for the information lost at the front end. Existing deep enhancement methods mostly focus on visual restoration and lack an end-to-end mapping that shares repair features with semantic decoding.

[0006] 4. Real-time challenges in engineering deployment: Many complex models rely on high-performance GPUs, making it difficult to meet navigation real-time and power consumption constraints directly on embedded AGV platforms. Lightweight and hierarchical inference strategies need to be designed to meet field deployment requirements.

[0007] Therefore, there is an urgent need for a processing method and system that can stably and efficiently identify QR codes under complex working conditions in order to improve the reliability and adaptability of AGV navigation. Summary of the Invention

[0008] The purpose of this invention is to provide a QR code image signal processing method and system for AGV navigation, thereby solving existing problems.

[0009] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: This invention relates to a method and system for processing QR code image signals for AGV navigation, comprising the following steps: S1. An original image containing a QR code is acquired using an image acquisition device installed on the AGV body; the image acquisition device is preferably a global shutter industrial camera with a resolution of not less than 640x480 to ensure the clarity and real-time performance of the image acquisition.

[0010] S2. The original image is input into the image restoration module, which performs noise suppression, contrast reconstruction, and local structure compensation on the original image, and outputs an enhanced image. The image restoration module adopts a convolutional neural network architecture containing an encoder and a decoder, preferably Residual U-Net, and adds multi-scale residual blocks between the encoder and decoder to enhance the high-frequency texture recovery capability. The network is constrained by a composite loss function during training. The composite loss function includes L1 / L2 pixel loss, perceptual loss (based on the intermediate layer features of the pre-trained VGG network, such as VGG Relu3_3, Relu4_3), structural or edge constraint loss (gradient difference, such as Laplacian), and adversarial loss (PatchGAN). The preferred weight ratio of each loss is _pixel:_perc:_struct:_adv=10:1:1:0.5, which can be adjusted according to the actual scene. The training samples of the image restoration module are synthesized in the following way: a semi-transparent oil stain layer is generated using Perlin noise (frequency range 0.5-8.0, amplitude 0.1-0.8). (Calibration) uses Gaussian and speckle noise superposition to simulate dust, and random polygon masks to simulate tearing and defects. Synthetic samples and real damaged-clear image pairs constitute a mixed training set to improve the model's adaptability to complex damaged scenes.

[0011] S3. The enhanced image is input into the QR code localization module to determine the corner positions or boundary regions of the QR code in the enhanced image. Based on the corner positions or boundary regions, the enhanced image is geometrically corrected to obtain a standardized QR code region image. The QR code localization module determines the corner positions by predicting keypoint coordinates through a regression network or by detecting keypoint heatmaps through a heatmap network. The regression network preferably uses a lightweight backbone such as MobileNetV3, outputting four corner coordinates (x1, y1...x4, y4). The loss is L2 regression loss supplemented by geometric consistency loss. The geometric correction is implemented through a spatial transformation network (STN), including calculating the homography matrix based on the predicted coordinates and using a GridGenerator and Bilinear Sampler to sample and correct the enhanced feature map. This process can be jointly trained with the keypoint regression network. The standardized QR code region image is a QR code feature image with uniform size, orientation, and scale, which facilitates subsequent decoding processing.

[0012] S4. The standardized QR code region image is input into the decoding module. The image is used to extract semantic features and predict sequences through a deep neural network to directly obtain the encoding information corresponding to the QR code. The decoding module first performs gridding processing on the standardized QR code region image, dividing it into NN grids (N is related to the QR code version). The local features of each grid are extracted and input into the Transformer encoder in the scanning order for global context modeling. The Self-Attention mechanism allows the model to refer to the feature distribution of the surrounding grids when predicting a certain blurred grid value, thereby achieving context compensation. Finally, a fully connected layer and softmax are used to directly output the bit stream or character sequence, and the confidence of each bit or character is output for subsequent policy decision-making.

[0013] S5. Output the encoded information to the AGV control system for AGV positioning or path planning; Steps S1 to S4 are executed periodically during AGV operation to achieve continuous recognition of ground QR codes; When the output confidence level is lower than the threshold, trigger repeated data collection, cloud-based high-precision recovery process or manual review process to ensure system safety and reliability.

[0014] A QR code recognition system for AGV navigation includes: an image acquisition module, an image restoration module, a QR code positioning and correction module, a decoding module, and an interface module, with each module connected in the order of data processing. The image acquisition module uses a global shutter industrial camera with a resolution of not less than 640x480. The image restoration module adopts a Residual U-Net architecture, trained based on a composite loss function, to achieve noise suppression, contrast reconstruction, and local structure compensation. The QR code positioning and correction module determines corner positions through a regression network or heatmap network and uses an STN layer for geometric correction. The decoding module uses a Transformer encoder to achieve semantic feature extraction and sequence prediction. The interface module outputs the encoded information and confidence level to the AGV control module and triggers corresponding processing strategies. The network models corresponding to each module undergo model quantization (such as INT8 quantization), structured pruning, knowledge distillation, or hierarchical inference processing to adapt to the computing resource limitations of embedded processors (such as NVIDIA Jetson, RK3568, RV1126), with a processing latency target of 30ms.

[0015] An automated guided vehicle (AGV) includes a vehicle body, an image acquisition device, a processor, and a memory. The memory stores a computer program, which executes the aforementioned QR code image signal processing method when running on the processor.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described QR code image signal processing method.

[0017] The present invention has the following beneficial effects: This invention utilizes a deep learning-based image restoration module to effectively repair complex industrial damage such as oil stains, dust, coverings, and defects, and restores QR code textures at the semantic level, overcoming the problem of traditional methods failing in severely damaged scenarios.

[0018] This invention uses keypoint regression or heatmap detection based on neural networks for QR code positioning. It does not rely on intact finder patterns and can still achieve stable and accurate geometric correction when finder patterns are worn or partially missing.

[0019] This invention couples the image restoration, localization correction and decoding processes at the feature level, and achieves semantic compensation between grids through Transformer-based context modeling, allowing the use of global information to infer locally blurred or missing modules, thus significantly improving the decoding success rate.

[0020] This invention achieves low latency processing of 30ms on embedded platforms such as NVIDIA Jetson and RK3568 by performing lightweight processing such as model quantization, pruning and distillation on the overall system, thus balancing high performance and low power consumption and meeting the real-time requirements of AGV navigation.

[0021] This invention introduces a confidence output and hierarchical decision-making mechanism. When the decoding confidence is low, it can trigger repeated data collection, high-precision cloud recovery, or manual review processes to form a closed-loop fault tolerance, ensuring the safety and stability of the AGV navigation process.

[0022] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 Traditional technology flowchart; Figure 2 This is a diagram of the overall system architecture of the present invention; Figure 3 This is a schematic diagram of training data synthesis. Figure 4 This is a network structure diagram. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0026] In the description of this invention, it should be understood that the terms "inside," "outside," "inside," etc., which indicate orientation or positional relationship, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the components or elements referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting this invention.

[0027] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, terms such as installation, setting, and connection should be interpreted broadly. For example, connection can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0028] Please see Figures 1-4 As shown, the present invention provides specific details of a QR code image signal processing system for AGV navigation: I. Hardware Environment and Deployment Examples 1. Acquisition end: A global shutter industrial camera with a resolution of 640x480 is used to ensure real-time image acquisition and clarity, and to avoid motion blur.

[0029] 2. Computing end: The embedded AI chip RV1126 is selected. Through optimization measures such as channel pruning, structured pruning, and INT8 quantization, the system processing latency is controlled within 30ms to meet the real-time requirements of AGV navigation.

[0030] II. Implementation of the Image Restoration Module 1. Data Synthesis: Generate a mixed training set as follows: a) Generate an oil smudge layer using Perlin noise (frequency range 0.58.0, amplitude calibrated to 0.10.8) and superimpose it onto the Alpha channel; b) Simulate dust using Gaussian and speckle noise superposition; c) Simulate tearing and defects using random polygon masks; d) Combine real damaged-clear QR code image pairs to construct a hybrid training set containing multiple damage types.

[0031] 2. Network Training: Residual U-Net was used as the network architecture, with multi-scale residual blocks added between the encoder and decoder. The loss function was L1 pixel loss + perceptual loss (VGG Relu3_3, Relu4_3) + structural edge loss (Laplacian) + PatchGAN adversarial loss, with the weight ratio set to _pixel:_perc:_struct:_adv=10:1:1:0.5. During training, the adaptive moment estimation (Adam) optimizer was used, with the initial learning rate set to 1e-4, which gradually decreased with each training iteration.

[0032] 3. Module Functions: Performs noise suppression, contrast reconstruction, and local structure compensation on the input raw image, focusing on processing low-contrast areas caused by oil stains, dust, water stains, or tire tracks, and outputs an enhanced image.

[0033] III. Implementation of Keypoint Regression and STN Correction 1. Keypoint Regression: MobileNetV3 is used as a lightweight backbone network to output the coordinates (x1, y1...x4, y4) of the four corner points of the QR code; the loss function adopts L2 regression loss, supplemented by geometric consistency loss, to ensure coordinate consistency on affine / perspective augmented data; during training, it is jointly trained with the image restoration module to improve localization accuracy.

[0034] 2. STN Correction: Calculate the homography matrix based on the predicted corner coordinates, generate a sampling grid using GridGenerator, and sample and correct the enhanced feature map using Bilinear Sampler to output a standardized QR code region image with uniform size, orientation, and scale.

[0035] IV. Semantic Decoding Implementation 1. Gridding: Divide the corrected and standardized QR code region image into N grids, where N is determined according to the QR code version (e.g., QR Code Version 1 is 2121 grids, Version 2 is 2525 grids, etc.).

[0036] 2. Sequence Modeling: Extract local features of each grid and input them into the Transformer encoder in a scanning order from left to right and from top to bottom. Establish contextual constraints between grids through the Self-Attention mechanism to achieve contextual compensation and improve the recognition accuracy of blurred grids.

[0037] 3. Output layer: The fully connected layer processes the output features of the Transformer encoder, directly outputting the bit stream or character sequence through the softmax activation function, and outputting the confidence score of each bit or character.

[0038] V. Post-processing and Decision Logic Implementation 1. Confidence level judgment: Set a confidence level threshold (e.g., 0.8). When the confidence level of the decoded output is higher than the threshold, the encoded information is directly output to the AGV control system for AGV positioning or path planning. 2. Low Confidence Handling: When the confidence level of the decoded output is lower than the threshold, one of the following processing strategies is triggered: a) Control the image acquisition device to repeatedly acquire the QR code image of the current area and re-execute the above processing procedure; b) Upload the damaged QR code image to the cloud server, process it using a high-precision recovery model, and then feed the encoded information back to the AGV control system; c) Trigger the manual verification process, where staff confirm the QR code information and input it into the AGV control system to ensure system safety and reliability.

[0039] VI. Experimental Verification To verify the technical effectiveness of this invention, QR code images collected in an actual factory environment were selected as test samples. On the embedded platform RV1126, the differences in decoding success rate between the method of this invention and traditional image processing methods were compared. The test results are shown in the table below: Test number Scene Description Type of damage Sample size Traditional methods readability Readability of this method T1 Light dust coverage Dust Coverage 100 92.0% 99.0% T2 Medium oil stains Oil stains 100 58.0% 95.0% T3 The locator part is missing Locator missing 100 40.0% 93.0% Experimental results show that the QR code readability of the method of the present invention is significantly higher than that of the traditional method under different damage scenarios. It can effectively improve the AGV's ability to recognize QR codes under complex working conditions, while meeting the real-time requirements of the embedded platform with a processing delay of 30ms.

[0040] In summary, this invention addresses many shortcomings of traditional QR code recognition algorithms in complex working conditions through generative network image restoration, spatial transformation network geometric correction, and end-to-end semantic decoding based on an attention mechanism, significantly improving the reliability and adaptability of AGV navigation. Furthermore, a lightweight strategy enables the system to be adapted to embedded AGV platforms, meeting the real-time and power consumption requirements of industrial environments, thus demonstrating broad application prospects and practical value.

[0041] In the description of this specification, references to terms such as "an embodiment," "an example," and "a specific example" indicate that a particular feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0042] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for processing QR code image signals for AGV navigation, characterized in that, Includes the following steps: S1. Acquire the original image containing the QR code using an image acquisition device installed on the AGV body; S2. Input the original image into the image restoration module, perform noise suppression, contrast reconstruction and local structure compensation processing, and output an enhanced image; S3. Input the enhanced image into the QR code positioning module to determine the corner position or boundary area of ​​the QR code in the image, and perform geometric correction to obtain a standardized QR code area image. S4. Input the standardized QR code area image into the decoding module, extract semantic features and predict sequences through a deep neural network, and output the encoding information and confidence level corresponding to the QR code. S5. Output the encoded information to the AGV control system for AGV positioning or path planning.

2. The QR code image signal processing method for AGV navigation according to claim 1, characterized in that, The image restoration module adopts a convolutional neural network structure that includes an encoder and a decoder, and is optimized using a composite loss function during training. The composite loss function includes at least pixel-level loss, perceptual loss, and structural consistency loss.

3. The QR code image signal processing method for AGV navigation according to claim 2, characterized in that, The training samples for the image restoration module are generated through synthesis, including: Perlin noise was used to simulate oil stains, Gaussian noise and speckle noise were superimposed to simulate dust, and random polygon masks were used to simulate image tearing or defects. Synthetic samples are combined with real damaged-clear image pairs to form a hybrid training set.

4. The QR code image signal processing method for AGV navigation according to claim 1, characterized in that, The QR code positioning module predicts corner coordinates through a regression network or detects corner positions through a heatmap network, and performs geometric correction through a spatial transformation network (STN) based on the position information to obtain a standardized QR code area image with uniform size, orientation, and scale.

5. A QR code image signal processing method for AGV navigation according to claim 1, characterized in that, The decoding module divides the standardized QR code region image into grids, extracts local features from each grid, and inputs them into the Transformer encoder for global context modeling. It achieves context compensation through a self-attention mechanism and finally outputs the encoded sequence and its confidence through a fully connected layer and softmax.

6. A QR code image signal processing method for AGV navigation according to claim 1, characterized in that, Steps S1 to S4 are executed periodically during AGV operation; when the output confidence level is lower than the set threshold, the process of repeated data collection, cloud-based high-precision recovery, or manual review is triggered.

7. A QR code image signal processing system for AGV navigation, characterized in that, include: Image acquisition module, used to acquire raw images containing QR codes; The image restoration module is used to perform noise suppression, contrast reconstruction, and local structure compensation on the original image, and output an enhanced image. The QR code positioning and correction module is used to detect QR code corners or boundaries, perform geometric correction, and output a standardized QR code area image. The decoding module is used to extract semantic features and perform sequence prediction through a deep neural network, and output encoded information and confidence level. The interface module is used to transmit coded information to the AGV control system and trigger corresponding processing strategies based on the confidence level.

8. A QR code image signal processing system for AGV navigation according to claim 7, characterized in that, The image restoration module adopts a Residual U-Net network structure, the QR code localization and correction module integrates a key point regression network and a spatial transformation network (STN), and the decoding module adopts a Transformer encoder structure. The system is deployed on an embedded processor, and the neural network models corresponding to each module undergo at least one lightweight processing of model quantization, structured pruning, knowledge distillation, or hierarchical inference to achieve a processing latency of no more than 30ms.

9. An automated guided vehicle (AGV), characterized in that, include: Vehicle body; An image acquisition device, installed on the vehicle body, is used to acquire images containing QR codes; processor; Memory, which stores computer programs; When the computer program is executed by the processor, it implements the QR code image signal processing method as described in any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the QR code image signal processing method as described in any one of claims 1 to 6.

Images