An image recognition system for AR glasses end

Abstract

The application discloses an image recognition system for AR glasses, which is deployed on AR glasses integrating a camera unit and an embedded AI chip and is composed of image acquisition, image preprocessing, edge enhancement small target detection and recognition, cross-frame steady tracking and task matching, and AR display processes. Firstly, images are acquired by the AR glasses camera unit and preprocessed to screen out qualified frames. Then, multi-scale features are extracted by using a lightweight backbone network, cross-scale feature fusion and edge enhancement mechanism to improve the positioning accuracy and recall rate of small targets, and the cross-frame continuous tracking of targets is realized in combination with a task matching mechanism. Finally, the recognition and task determination results are displayed in real time on the AR glasses. Compared with the prior art, the application can realize high-precision, low-delay and high-stability recognition of industrial small targets in offline and other complex environments, significantly reduces the dependence on cloud computing, and is suitable for various industrial application scenarios such as device monitoring and automatic inspection.

Description

Technical Field

[0001] This invention belongs to the field of industrial visual inspection and augmented reality (AR) wearable device technology, specifically relating to an image recognition system for AR glasses. Background Technology

[0002] Image recognition for AR glasses is a crucial technology in intelligent industrial inspection, aiming to enable real-time perception, recognition, and result prompts for targets such as instruments, indicator lights, and buttons in inspection scenarios. While the overall performance of image recognition has improved with the continuous development of augmented reality and edge intelligence technologies, significant challenges remain in AR glasses image recognition within industrial inspection scenarios. The following issues are common in industrial images: First, targets such as instrument scales, indicator lights, and buttons occupy a small area and are minute in scale, representing typical small targets that are prone to weak features and difficulty in localization. Second, industrial environments are often plagued by glare, occlusion, cluttered backgrounds, and uneven brightness, easily causing blurred target edges, missing texture information, and confusion between targets and background. Third, AR glasses are also affected by factors such as changing viewing angles, frame jitter, and limited edge resources during mobile inspections, leading to fluctuations in continuous recognition results and difficulty in maintaining stability. These factors collectively contribute to a decline in the accuracy and stability of AR glasses image recognition in industrial inspection scenarios.

[0003] To address the challenges posed by complex industrial environments, existing technologies primarily employ two approaches: First, cloud-based recognition and processing. This involves AR glasses capturing on-site images or video streams and uploading them to a cloud server. The target detection and recognition model on the server then analyzes and processes the data, before transmitting the results back to the terminal for display. This method achieves high recognition accuracy under favorable network conditions but suffers from high dependence on network bandwidth and transmission stability, as well as significant communication latency. Second, edge-based lightweight recognition. This involves trimming, compressing, or lightweighting target detection networks such as YOLOv5 and YOLOv8 and deploying them on edge devices like smartphones, tablets, or AR glasses for local real-time detection. Additionally, some technologies improve the visibility of small target areas in inspection images through image enhancement or super-resolution reconstruction to enhance detection performance. While the methods mentioned above improve edge processing capabilities or image quality to some extent, most general lightweight models are designed for training in ordinary scenarios. When faced with small targets such as instrument scales, indicator lights, and buttons in industrial settings, they are still prone to issues such as missed detections, positioning errors, and decreased recognition performance in complex backgrounds. Image enhancement or super-resolution methods typically have high computational overhead and mainly focus on improving the quality of a single frame, making it difficult to meet the requirements of real-time processing and continuous stable recognition on AR glasses.

[0004] To address the aforementioned shortcomings, especially in complex industrial environments, targets such as instrument scales, indicator lights, and buttons are often small in size and frequently subject to interference from reflections, obstructions, cluttered backgrounds, and brightness variations. This leads to problems such as indistinct features, blurred edges, and inaccurate positioning in captured images. Improving the recognition accuracy of small industrial targets in complex scenarios has become a key challenge. Currently, there is a lack of effective methods to enhance the feature representation of small targets under complex background conditions and improve their detection accuracy and recognition stability. Therefore, it is necessary to propose a new technical solution to address the issues of missed detections, false detections, and positioning errors of small targets in AR glasses image recognition within complex industrial environments. Summary of the Invention

[0005] To address the aforementioned issues, this invention discloses an image recognition system for AR glasses. By integrating an edge-enhanced small target detection and recognition system into the AR glasses and combining it with cross-frame steady-state tracking and task matching mechanisms, it achieves high-precision recognition, low-power real-time detection, and task-level judgment of small targets. This reduces reliance on the cloud during industrial inspections and improves the system's intelligence and automation levels, meeting the inspection needs of various industrial scenarios.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] An image recognition system for AR glasses includes the following steps:

[0008] Step 1: Image Acquisition. The camera unit at the AR glasses acquires images or video streams of the industrial inspection scene at a preset frame rate;

[0009] Step 2: Image preprocessing. Geometric correction, brightness normalization, and noise suppression are performed on the acquired images. Image quality is then screened based on a multi-index fusion sharpness evaluation, and only images meeting the required quality are sent to the recognition module.

[0010] Step 3: Edge-enhanced small target detection and recognition. Image features are extracted based on a lightweight backbone network, and the feature response of small target regions is enhanced and background noise is suppressed through cross-scale feature fusion and edge enhancement mechanisms. The small target detection head outputs small target detection results including category, bounding box coordinates, and confidence score.

[0011] Step 4: Cross-frame steady-state tracking and task matching. Based on the small target detection results, cross-frame data association, target state prediction, and steady-state determination are performed, and inspection task matching is completed according to the task configuration file;

[0012] Step 5: AR Display. Upon completion of the task, keyframe images of the corresponding steady-state trajectory are selected, compressed, and output to generate an inspection record containing target category, spatial coordinates, steady-state duration, and task judgment information, which is then displayed in real time on the AR glasses.

[0013] As a further supplement to the present invention, the image acquisition in step one is specifically as follows:

[0014] The AR glasses' built-in high-resolution camera captures images or video streams of industrial inspection scenes in real time. Under different lighting and environmental conditions, it automatically adjusts exposure and focus. The captured images are continuously acquired at a set frame rate. After being transmitted to the processing unit in real time, a sharpness detection algorithm monitors the quality to ensure that the images meet the requirements of subsequent processing.

[0015] As a further supplement to the present invention, the image preprocessing in step two is as follows:

[0016] First, the acquired images are analyzed for sharpness using Laplacian variance and Tenengrad gradient energy. Laplacian variance measures image sharpness, and the formula is:

[0017] (1-1);

[0018] in It is the grayscale value of a pixel in the image. It is the average gray value of the image. and These are the width and height of the image. A larger Laplacian variance value indicates a sharper image. The Tenengrad gradient energy is used to calculate the edge information of the image, evaluate its sharpness, and determine whether the image is clear. The formula is:

[0019] (1-2);

[0020] in and These are the gradient values ​​of the image in the x-axis and y-axis directions, respectively.

[0021] Due to factors such as camera angle and lens distortion, images may exhibit distortion or falsification. Therefore, geometric correction is necessary, using perspective transformation and distortion correction to correct geometric errors in the image. Perspective transformation typically uses a homography matrix to correct geometric distortion; the transformation formula is as follows:

[0022] (1-3);

[0023] in These are the corrected image coordinates. Here are the original image coordinates, and H is a 3×3 homography matrix representing the geometric relationship between the two images. Distortion correction is performed using the camera's intrinsic and extrinsic reference images, using the following formula:

[0024] (1-4);

[0025] in and These are the distorted image pixels. and It is the distortion coefficient. These are the corrected image pixels.

[0026] Different lighting conditions can lead to uneven brightness distribution in images. Therefore, it's necessary to normalize the image's brightness, mapping pixel values ​​to a uniform brightness range to reduce the impact of lighting variations. Common methods include linear normalization and histogram equalization. The formula for linear normalization is:

[0027] (1-5);

[0028] in, These are pixel values ​​in the image. and These are the minimum and maximum values ​​of the image, respectively. These are the normalized pixel values. Histogram equalization adjusts the contrast of an image to make the brightness distribution more uniform. The formula is:

[0029] (1-6);

[0030] in It is a pixel The cumulative probability, It is the range of pixel values ​​in the image. It is a pixel frequency, It is the number of pixels in the image.

[0031] Images in industrial environments are often affected by noise, especially in low-light conditions. Therefore, noise suppression is performed to remove high-frequency noise while preserving the most important image features. Noise suppression methods include mean filtering and Gaussian filtering. The mean filtering formula is as follows:

[0032] (1-7);

[0033] in These are the filtered image pixels. is the value of the neighboring pixels in the original image, and K is the radius of the filter. The weights of the Gaussian filter are calculated using the Gaussian distribution function, as shown in the formula:

[0034] (1-8);

[0035] in, It is a Gaussian kernel. It is the standard deviation, which controls the range and intensity of the filter.

[0036] As a further supplement to the present invention, the edge-enhanced small target detection and recognition in step three is specifically as follows:

[0037] In complex contexts, traditional small target detection methods suffer from low accuracy due to limited receptive field and insufficient feature information. To address this issue, an innovative edge-enhanced small target detection and recognition algorithm is proposed. By improving the convolutional neural network architecture and combining Deformed Large Kernel Attention (D-LKA), Cross-Scale Feature Fusion (CCFF), and Efficient Cross-Intersection Over Union (EIOU), the accuracy and real-time performance of small target detection are significantly improved.

[0038] The backbone network uses the lightweight MobileNetV3, with optimized inverse residual structure and the introduction of a hard Swish activation function. This reduces computational cost while maintaining high-efficiency feature extraction capabilities. The hard Swish activation function formula is as follows:

[0039] (1-9);

[0040] in, This represents the feature value input to the hard Swish activation function, where the feature value is the intermediate activation value output by the current network layer. This indicates that the input is restricted to a range. The linear rectifier function within.

[0041] Compared to the traditional Swish activation function, Hard Swish significantly improves computational efficiency by simplifying the calculation process, making it particularly suitable for resource-constrained devices. The network neck structure integrates a cross-scale feature fusion (CCFF) algorithm, enhancing the detection accuracy for small targets. By fusing multi-scale features, CCFF effectively preserves detailed information in low-resolution features and combines it with high-level semantic information to enhance the recognition ability of small targets. Assuming the feature map output from the backbone network is... , , These represent feature representations at different scales; by adjusting the channel dimensions through 1×1 convolution, the corresponding channel-aligned feature maps are obtained. , , The formula is:

[0042] (1-10);

[0043] in, , , Representing feature maps respectively , , Channel adjustment result after 1×1 convolution.

[0044] Then, the feature maps are weighted and fused to obtain the final feature representation:

[0045] (1-11);

[0046] in , , These are the learned weighting coefficients, ensuring that low-level and high-level features are effectively combined, thus enhancing the detection capability of small targets.

[0047] When dealing with complex backgrounds, conventional convolutional operations have a small receptive field, making it difficult to capture global contextual information. The Deformed Large Kernel Attention (D-LKA) mechanism expands the receptive field through depthwise convolution to extract local features and depthwise dilated convolution, adjusting channel dimensions to further enhance the detection capability for small objects. This allows the network to adaptively focus on important regions in the image. The depthwise convolution formula is:

[0048] (1-12);

[0049] Where d is the dilation rate, controlling the size of the region covered by the convolution kernel. The formula for depthwise dilated convolution is:

[0050] (1-13);

[0051] in d is the size of the convolution kernel, and d is the dilation rate.

[0052] To improve positioning accuracy, the efficient intersection-union ratio (EIOU) loss is used instead of the traditional CIOU loss. EIOU loss significantly improves positioning accuracy by explicitly modeling the width and height deviations between the predicted bounding box and the ground truth bounding box, as well as the width and height differences of the minimum bounding rectangle. The formula for EIOU loss is:

[0053] (1-14);

[0054] in It is a comparison of losses. It is the center point distance loss. It is the aspect ratio loss. By decoupling the loss calculations for width and height, EIOU can accelerate the training process and improve positioning accuracy.

[0055] As a further supplement to the present invention, the cross-frame steady-state tracking and task matching in step four are specifically as follows:

[0056] After obtaining the single-frame detection results, to ensure the consistent representation of the target in a continuous image sequence, a cross-frame tracking mechanism based on a state-space model is constructed to generate the target's steady-state trajectory, and matching and judgment for the inspection task are then performed based on this trajectory. First, basic state variables are constructed for the detection boxes, including the coordinates of the bounding box center point. Width and height Including the rate of change of position, a simplified linear motion model is used to predict the target's position in the next frame. The target's state vector can be represented as:

[0057] (1-15);

[0058] The state of the target in the next frame is predicted using the state transition formula:

[0059] (1-16);

[0060] in, Indicates the first The state vector of the frame target. , Indicates the coordinates of the target's center point. , This indicates the width and height of the target bounding box. , This represents the rate of change of the target center point's position in the horizontal and vertical directions. This indicates the prediction result for the target state in the next frame. This is the state transition matrix, used to describe the state evolution relationship where the target center point moves at approximately a constant velocity and the width and height of the bounding box remain relatively stable.

[0061] To match the predicted bounding box with the detected bounding box in the current frame, overlap and spatial distance are used as matching criteria. With detection box The intersection-union ratio (IoU) is calculated as follows:

[0062] (1-17);

[0063] in, This represents the area of ​​the intersection region between the predicted bounding box and the detected bounding box. This represents the area of ​​the union region of the predicted bounding box and the detected bounding box.

[0064] To compensate for the shortcomings of IoU in small target scenarios, the center distance is further calculated and defined as:

[0065] (1-18);

[0066] in, This represents the Euclidean distance between the center of the predicted bounding box and the center of the detected bounding box. Represents the prediction box The center coordinates, Represents the detection box The center coordinates.

[0067] Further construct the joint cost function:

[0068] (1-19);

[0069] in, This represents the matching cost between the predicted bounding box and the detected bounding box. and Matching weight parameters are used to adjust the relative influence of the intersection-union ratio (IU) and center distance (CUM) terms during the matching process. By minimizing the cost function and employing a Hungarian matching strategy, globally optimal association of multiple targets in consecutive frames is achieved.

[0070] After cross-frame association, to avoid interference from detection jitter on the tracking trajectory, a steady-state determination mechanism is introduced. This mechanism is used when the target's position change over several consecutive frames satisfies the following conditions:

[0071] (1-20);

[0072] And the detection confidence level remains at the threshold. When the above conditions are met, the target is considered to have entered a steady-state tracking state. Among these, Indicates the target is in the first place. Frame and the The change in center position between frames Indicates the target is in the first place. The center coordinates in the frame Indicates the target is in the first place. The center coordinates in the frame Indicates the threshold for position change. This represents the detection confidence threshold.

[0073] Based on the inspection task configuration file, the system determines the target category, spatial location, and steady-state duration. When the detected target category matches the task requirements, and the target maintains a steady state within the specified area for more than a predetermined threshold, the system will proceed. When the conditions for the inspection task are met, it is considered that the task has been completed.

[0074] As a further supplement to the present invention, the AR display in step five is as follows:

[0075] After the target completes steady-state tracking and task matching, the system automatically selects high-resolution and representative keyframe images based on the steady-state interval and displays the target category, spatial coordinates, steady-state duration, and task judgment information in real time on the AR glasses, providing intuitive prompts for on-site inspection personnel.

[0076] The beneficial effects of this invention are as follows:

[0077] This invention discloses an image recognition system for AR glasses, which introduces an edge-enhanced small target detection and recognition algorithm. This algorithm eliminates the need for additional complex feature enhancement networks. Multi-scale feature fusion and a small target enhancement branch effectively improve the recall rate and localization accuracy of small targets in low-light, reflective, and occluded environments. Furthermore, by combining cross-frame steady-state tracking and task matching mechanisms, the system further enhances the stability of target detection and the accuracy of task completion determination. Verification shows that this system not only improves small target detection performance but also reduces reliance on cloud computing, optimizing the utilization of computational resources. It is particularly suitable for industrial inspection applications on resource-constrained devices such as AR glasses. Attached Figure Description

[0078] Figure 1 This is a flowchart of an image recognition system for AR glasses proposed in this invention;

[0079] Figure 2 This is a structural diagram of the edge-enhanced small target detection and recognition module proposed in this invention. Detailed Implementation

[0080] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention.

[0081] like Figure 1 As shown, this invention proposes an image recognition system for AR glasses. The system comprises image acquisition, image preprocessing, edge enhancement and small target detection and recognition, cross-frame steady-state tracking and task matching, and AR display, including the following steps:

[0082] Step 1: The high-resolution camera built into the AR glasses captures images or video streams in real time from the industrial inspection scene, and sets the frame rate. Continuous output of raw image sequence

[0083] (2-1);

[0084] in, Indicates the first Frame-by-frame image acquisition This indicates the total number of frames captured. During the inspection process, the camera continuously samples the equipment area within its current field of view, enabling the continuous acquisition of small industrial targets such as instruments, indicator lights, and buttons. The acquired images are transmitted to the processing unit in real-time according to time sequence, serving as the raw input data for subsequent image preprocessing, small target detection and recognition, and cross-frame steady-state tracking.

[0085] Step Two: First, perform sharpness analysis on the acquired images, using Laplacian variance and Tenengrad gradient energy for evaluation. Laplacian variance is used to measure image sharpness, and the formula is:

[0086] (2-2);

[0087] in It is the grayscale value of a pixel in the image. It is the average gray value of the image. and These are the width and height of the image. A larger Laplacian variance value indicates a sharper image. The Tenengrad gradient energy is used to calculate the edge information of the image, evaluate its sharpness, and determine whether the image is clear. The formula is:

[0088] (2-3);

[0089] in and These are the gradient values ​​of the image in the x-axis and y-axis directions, respectively.

[0090] Due to factors such as camera angle and lens distortion, images may exhibit distortion or falsification. Therefore, geometric correction is necessary, using perspective transformation and distortion correction to correct geometric errors in the image. Perspective transformation typically uses a homography matrix to correct geometric distortion; the transformation formula is as follows:

[0091] (2-4);

[0092] in These are the corrected image coordinates. Here are the original image coordinates, and H is a 3×3 homography matrix representing the geometric relationship between the two images. Distortion correction is performed using the camera's intrinsic and extrinsic reference images, using the following formula:

[0093] (2-5);

[0094] in and These are the distorted image pixels. and It is the distortion coefficient. These are the corrected image pixels.

[0095] Different lighting conditions can lead to uneven brightness distribution in images. Therefore, it's necessary to normalize the image's brightness, mapping pixel values ​​to a uniform brightness range to reduce the impact of lighting variations. Common methods include linear normalization and histogram equalization. The formula for linear normalization is:

[0096] (2-6);

[0097] in, These are pixel values ​​in the image. and These are the minimum and maximum values ​​of the image, respectively. These are the normalized pixel values. Histogram equalization adjusts the contrast of an image to make the brightness distribution more uniform. The formula is:

[0098] (2-7);

[0099] in It is a pixel The cumulative probability, It is the range of pixel values ​​in the image. It is a pixel frequency, It is the number of pixels in the image.

[0100] Images in industrial environments are often affected by noise, especially in low-light conditions. Therefore, noise suppression is performed to remove high-frequency noise while preserving the most important image features. Noise suppression methods include mean filtering and Gaussian filtering. The mean filtering formula is as follows:

[0101] (2-8);

[0102] in These are the filtered image pixels. is the value of the neighboring pixels in the original image, and K is the radius of the filter. The weights of the Gaussian filter are calculated using the Gaussian distribution function, as shown in the formula:

[0103] (2-9);

[0104] Where G(x,y) is the Gaussian kernel, It is the standard deviation, which controls the range and intensity of the filter.

[0105] Step 3: Figure 2This is a module structure diagram of edge-enhanced small target detection and recognition designed in this invention. It is based on a lightweight backbone network and combines cross-scale feature fusion and deformable large kernel attention mechanism to perform feature extraction and target detection on preprocessed images.

[0106] The backbone network uses the lightweight MobileNetV3, with optimized inverse residual structure and the introduction of a hard Swish activation function. This reduces computational cost while maintaining high-efficiency feature extraction capabilities. The hard Swish activation function formula is as follows:

[0107] (2-10);

[0108] in, This represents the feature value input to the hard Swish activation function, where the feature value is the intermediate activation value output by the current network layer. This indicates that the input is restricted to a range. The linear rectifier function within.

[0109] Compared to the traditional Swish activation function, Hard Swish significantly improves computational efficiency by simplifying the calculation process, making it particularly suitable for resource-constrained devices. The network neck structure integrates a cross-scale feature fusion (CCFF) algorithm, enhancing the detection accuracy for small targets. By fusing multi-scale features, CCFF effectively preserves detailed information in low-resolution features and combines it with high-level semantic information to enhance the recognition ability of small targets. Assuming the feature map output from the backbone network is... , , These represent feature representations at different scales; by adjusting the channel dimensions through 1×1 convolution, the corresponding channel-aligned feature maps are obtained. , , The formula is:

[0110] (2-11);

[0111] in, , , Representing feature maps respectively , , Channel adjustment result after 1×1 convolution.

[0112] Then, the feature maps are weighted and fused to obtain the final feature representation:

[0113] (2-12);

[0114] in , , These are the learned weighting coefficients, ensuring that low-level and high-level features are effectively combined, thus enhancing the detection capability of small targets.

[0115] When dealing with complex backgrounds, conventional convolutional operations have a small receptive field, making it difficult to capture global contextual information. The Deformed Large Kernel Attention (D-LKA) mechanism expands the receptive field through depthwise convolution to extract local features and depthwise dilated convolution, adjusting channel dimensions to further enhance the detection capability for small objects. This allows the network to adaptively focus on important regions in the image. The depthwise convolution formula is:

[0116] (2-13);

[0117] in The dilation rate controls the size of the region covered by the convolution kernel. The formula for depthwise dilated convolution is:

[0118] (2-14);

[0119] in It is the size of the convolution kernel. It is the expansion rate.

[0120] To improve positioning accuracy, the efficient intersection-union ratio (EIOU) loss is used instead of the traditional CIOU loss. EIOU loss significantly improves positioning accuracy by explicitly modeling the width and height deviations between the predicted bounding box and the ground truth bounding box, as well as the width and height differences of the minimum bounding rectangle. The formula for EIOU loss is:

[0121] (2-15);

[0122] in It is a comparison of losses. It is the center point distance loss. It is the aspect ratio loss. By decoupling the loss calculations for width and height, EIOU can accelerate the training process and improve positioning accuracy.

[0123] Step 4: After obtaining the single-frame detection results, to ensure the consistent representation of the target in the continuous image sequence, a cross-frame tracking mechanism based on a state-space model is constructed to generate the target's steady-state trajectory, and matching and judgment for the inspection task are performed based on this. First, basic state variables are constructed for the detection boxes, including the coordinates of the bounding box center point. Width and height Including the rate of change of position, a simplified linear motion model is used to predict the target's position in the next frame. The target's state vector can be represented as:

[0124] (2-16);

[0125] The state of the target in the next frame is predicted using the state transition formula:

[0126] (2-17);

[0127] in, Indicates the first The state vector of the frame target. , Indicates the coordinates of the target's center point. , This indicates the width and height of the target bounding box. , This represents the rate of change of the target center point's position in the horizontal and vertical directions. This indicates the prediction result for the target state in the next frame. This is the state transition matrix, used to describe the state evolution relationship where the target center point moves at approximately a constant velocity and the width and height of the bounding box remain relatively stable.

[0128] To match the predicted bounding box with the detected bounding box in the current frame, overlap and spatial distance are used as matching criteria. With detection box The intersection-union ratio (IoU) is calculated as follows:

[0129] (2-18);

[0130] in, This represents the area of ​​the intersection region between the predicted bounding box and the detected bounding box. This represents the area of ​​the union region of the predicted bounding box and the detected bounding box.

[0131] To compensate for the shortcomings of IoU in small target scenarios, the center distance is further calculated and defined as:

[0132] (2-19);

[0133] in, This represents the Euclidean distance between the center of the predicted bounding box and the center of the detected bounding box. Represents the prediction box The center coordinates, Represents the detection box The center coordinates.

[0134] Further construct the joint cost function:

[0135] (2-20);

[0136] in, This represents the matching cost between the predicted bounding box and the detected bounding box. and Matching weight parameters are used to adjust the relative influence of the intersection-union ratio (IU) and center distance (CUM) terms during the matching process. By minimizing the cost function and employing a Hungarian matching strategy, globally optimal association of multiple targets in consecutive frames is achieved.

[0137] After cross-frame association, to avoid interference from detection jitter on the tracking trajectory, a steady-state determination mechanism is introduced. This mechanism is used when the target's position change over several consecutive frames satisfies the following conditions:

[0138] (2-21);

[0139] And the detection confidence level remains at the threshold. When the above conditions are met, the target is considered to have entered a steady-state tracking state. Among these, Indicates the target is in the first place. Frame and the The change in center position between frames Indicates the target is in the first place. The center coordinates in the frame Indicates the target is in the first place. The center coordinates in the frame Indicates the threshold for position change. This represents the detection confidence threshold.

[0140] Based on the inspection task configuration file, the system determines the target category, spatial location, and steady-state duration. When the detected target category matches the task requirements, and the target maintains a steady state within the specified area for more than a predetermined threshold, the system will proceed. When the conditions for the inspection task are met, it is considered that the task has been completed.

[0141] Step 5: After the target completes steady-state tracking and task matching, the system automatically selects high-resolution and representative keyframe images based on the corresponding steady-state interval, and compresses and outputs these keyframe images. Subsequently, the keyframe images, target category, spatial coordinates, steady-state duration, and task determination information are correlated to generate an inspection record, which can be represented as:

[0142] (2-22);

[0143] in, This indicates the inspection record. Represents the keyframe image. Indicates target category information, Represents the target spatial coordinate information. Indicates the duration of steady state. This indicates the task determination information. The system displays the inspection records in real time on the AR glasses to provide intuitive prompts to on-site inspection personnel.

[0144] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. An image recognition system for AR glasses, characterized in that, Includes the following steps: Step 1: Image Acquisition: The camera unit at the AR glasses acquires images or video streams of the industrial inspection scene at a preset frame rate; Step 2: Image preprocessing: Geometric correction, brightness normalization and noise suppression are performed on the acquired images, and the image quality is screened based on multi-index fusion sharpness evaluation. Only images that meet the quality requirements are sent to the recognition module. Step 3: Edge-enhanced small target detection and recognition: Image features are extracted based on a lightweight backbone network, and the feature response of small target regions is enhanced and background noise is suppressed through cross-scale feature fusion and edge enhancement mechanisms. The small target detection head outputs small target detection results including category, bounding box coordinates and confidence. Step 4: Cross-frame steady-state tracking and task matching: Based on the small target detection results, perform cross-frame data association, target state prediction and steady-state determination, and complete the inspection task matching according to the task configuration file; Step 5: AR Display: Upon completion of the task, select the keyframe image of the corresponding steady-state trajectory, compress and output it to generate an inspection record containing target category, spatial coordinates, steady-state duration and task judgment information, and display it in real time on the AR glasses.

2. The system according to claim 1, characterized in that, In step three, the lightweight backbone network uses an inverted residual structure and a hard Swish activation function to extract multi-scale features. It then uses a cross-scale feature fusion unit to perform weighted fusion of features at adjacent scales, preserving shallow edge details and introducing high-level semantic information. The edge enhancement mechanism is based on a deformable large kernel attention mechanism to enhance features in small target regions and suppress complex background responses.

3. The system according to claim 2, characterized in that, The edge enhancement mechanism employs a deformable large kernel attention mechanism, which expands the local receptive field of small target regions through depthwise convolution, where the depthwise convolution satisfies: ； in, This represents the effective receptive field of depthwise convolution. The dilation rate controls the size of the region covered by the convolution kernel.

4. The system according to claim 2, characterized in that, The cross-scale feature fusion unit uses learnable weighted coefficients to combine shallow edge features and deep semantic features to improve the feature discrimination ability of small target detection.

5. The system according to claim 2, characterized in that, The small target detection head in step three includes a dedicated prediction branch for small targets, an efficient intersection-over-union (IoU) bounding box regression loss, and a small target reweighting strategy. It performs small target detection on high-resolution feature maps, and the bounding box regression loss satisfies the following: ； in, To compare the losses, For the center point distance loss, The aspect ratio loss is used to improve the positioning accuracy and recall rate of small targets by decoupling the loss calculation of width and height.

6. The system according to claim 1, characterized in that, The cross-frame steady-state tracking and task matching in step four includes: cross-frame data association based on the current frame detection results and historical frame trajectory results; target state prediction based on target position, bounding box size and motion trend; target matching by combining the prediction results with the current frame detection results; and inspection task determination based on the target category, spatial position and continuous state according to the task configuration file.

7. The system according to claim 6, characterized in that, In step four, the target state prediction is expressed as: ； in, Indicates the first The state vector of the frame target. , Indicates the coordinates of the target's center point. , This indicates the width and height of the target bounding box. , This represents the rate of change of the target's center point in the horizontal and vertical directions. The system completes cross-frame matching based on the overlap between the predicted box and the detection box and the center distance. When the target's position change is less than a preset threshold, the detection confidence is higher than a preset threshold, and the task area constraint is met, the system determines that the target has entered a steady state and completes the task matching.

Images