An object size precision measurement method and system based on a DU-YOLO architecture

Abstract

This invention discloses a method and system for precise measurement of object dimensions based on the DU-YOLO architecture. The method includes: Step S1. Acquiring an image dataset of the object and performing annotation and enhancement processing on the acquired image dataset to obtain an enhanced dataset; Step S2. Constructing an improved DU-YOLO detection model, wherein the improved DU-YOLO detection model uses the YOLO13-Pose network as the baseline model, introduces a dynamically evolving aggregation architecture DEA-Arch structure to replace the original backbone network in the backbone network, and introduces a practical balanced efficiency detection head UBEH structure to replace the original detection head in the detection head; Step S3. Training the improved DU-YOLO keypoint detection model using the enhanced dataset from Step S1 to obtain a trained keypoint detection model; Step S4. Inputting the image of the object to be measured into the trained keypoint detection model for keypoint detection to obtain keypoint detection and precise dimension measurement results.

Description

Technical Field

[0001] This invention relates to the field of machine vision and target detection technology, and in particular to a method and system for precise measurement of object size based on the DU-YOLO architecture. Background Technology

[0002] In modern electronics manufacturing, dimensional measurement of components and products is a crucial step in ensuring product quality. Traditional manual visual inspection is inefficient, prone to missed or incorrect inspections, and cannot meet the demands of high-precision, high-efficiency production. Conventional machine vision methods are significantly affected by environmental factors such as lighting and occlusion, resulting in poor robustness. With the rapid development of RGB-D data acquisition equipment and deep learning technology, machine vision-based dimensional measurement has become the mainstream trend in the industry. However, while existing keypoint detection models based on YOLO13-Pose can achieve target detection and keypoint localization in a single stage, they still have shortcomings in the following aspects: (1) The backbone network has a large amount of computation and many parameters, which is not conducive to the deployment of edge devices; (2) The positioning accuracy of key points of small targets is insufficient, and the measurement error is large in complex industrial scenarios.

[0003] Therefore, in view of the above-mentioned prior art, the present invention provides a method and system for precise measurement of object size based on the DU-YOLO architecture. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and system for precise measurement of object dimensions based on the DU-YOLO architecture. This method can balance real-time performance with high-precision lightweight key point detection to meet the actual needs of intelligent dimension measurement in industrial production lines.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for precise measurement of object dimensions based on the DU-YOLO architecture, comprising: Step S1. Obtain the image dataset of the object, and perform annotation and augmentation processing on the obtained image dataset to obtain the augmented dataset; Step S2. Construct an improved DU-YOLO detection model, wherein the improved DU-YOLO detection model is based on the YOLO13-Pose network as the baseline model, and introduces the dynamic evolution aggregation architecture DEA-Arch structure in the backbone network to replace the original backbone network, and introduces the practical balanced efficiency detection head UBEH structure in the detection head to replace the original detection head. Step S3. Use the augmented dataset from step S1 to train the improved DU-YOLO keypoint detection model to obtain the trained keypoint detection model; Step S4. Input the image of the object to be tested into the trained key point detection model to perform key point detection and obtain key point detection and precise size measurement results.

[0006] Furthermore, the dynamically evolving aggregation architecture DEA-Arch structure in step S2 includes: The dynamic evolution unit is used to reduce the dimensionality of the input features during the training phase. It extracts features by combining a multi-branch structure with channel rearrangement and merges them into a single-path convolution during the inference phase through structural reparameterization. The single-step aggregation mechanism is used to aggregate features output by dynamic evolution units to achieve multi-scale feature fusion and output.

[0007] Furthermore, the multi-branch structure in the dynamic evolution unit includes at least a 2×2 convolution branch, a 1×1 convolution branch, and a residual connection branch.

[0008] Furthermore, the UBEH structure for the practical balance efficiency detection head in step S2 includes: The task decoupling module is built on the Transformer decoder architecture and has separate first output interface, second output interface and third output interface, which are used to output target classification information, bounding box regression information and key point localization information, respectively. The adaptive deformation module has a first input end and a second input end, which are respectively connected to the internal feature nodes in the task decoupling module that correspond to the bounding box regression information and the key point localization information. The feature adaptation module, connected to the adaptive deformation module, is used to provide the adaptive deformation module with learnable offset parameters and modulation mask parameters.

[0009] Furthermore, the UBEH practical balance efficiency detection head structure also includes: The pyramid-shaped multi-scale feature fusion module has its input end connected to the output end of the backbone network. It is used to receive and fuse feature maps from different scales of the backbone network. Its multiple output ends are respectively connected to the task input ends of the task decoupling layer. It is used to perform weighted concatenation and context information enhancement processing on the feature layers of different scales output by the backbone network, and input the processed features to the task decoupling layer.

[0010] Furthermore, step S1 specifically includes: Images containing objects and their corresponding depth maps are acquired. The acquired images undergo a first data augmentation process. Bounding boxes and keypoint annotations are added to the first augmented images. The annotated images undergo a second augmentation process to obtain the augmented dataset.

[0011] Furthermore, the first round of enhancement processing includes at least one of rotation, flipping, brightness adjustment, and random occlusion; the second enhancement processing includes at least one of blurring, adding noise, color jittering, and random erasure.

[0012] Furthermore, the key point detection results in step S4 include at least the bounding box coordinates of the object, the category confidence score, and the pixel coordinates of the key points in the image of the object to be tested.

[0013] Furthermore, the precise size measurement results in step S4 are calculated based on the pixel coordinates of key points, camera intrinsic parameters, and depth map information.

[0014] Correspondingly, a precision measurement system for object dimensions based on the DU-YOLO architecture is also provided, including: The acquisition module is used to acquire image datasets of objects and perform annotation and augmentation processing on the acquired image datasets to obtain augmented datasets; The building module is used to build an improved DU-YOLO detection model. The improved DU-YOLO detection model is based on the YOLO13-Pose network. The original backbone network is replaced by a dynamically evolving aggregation architecture DEA-Arch structure, and the original detection head is replaced by a practical balanced efficiency detection head UBEH structure. The training module is used to train the improved DU-YOLO keypoint detection model using an augmented dataset to obtain a trained keypoint detection model. The detection module is used to input the image of the object to be measured into the trained key point detection model to perform key point detection and obtain key point detection and precise size measurement results.

[0015] Compared with existing technologies, the precise object size measurement method constructed in this invention introduces a DEA-Arch structure into the backbone network, significantly reducing the number of parameters and computational complexity, while improving the feature representation capability of small targets. A UBEH structure is introduced into the detection head, and through complete task decoupling, adaptive deformation modules, feature adaptive modules, and pyramid-style multi-scale feature fusion modules, the localization accuracy and robustness of key points in complex scenes are greatly improved. The precise object size measurement method constructed in this invention achieves the best balance between detection accuracy and real-time performance. In complex industrial scenarios such as small targets, heavy occlusion, varying lighting, and dense arrangements, key point localization is more accurate, size measurement errors are smaller, and measurement accuracy reaches the sub-pixel level. It is particularly suitable for online intelligent detection of object sizes with stringent accuracy requirements, possessing excellent practical application value and promising prospects for widespread application. Attached Figure Description

[0016] Figure 1This is a flowchart of a method for precise measurement of object dimensions based on the DU-YOLO architecture, provided in Embodiment 1. Figure 2 This is a schematic diagram of the overall DU-YOLO network structure provided in Example 1. Figure 3 This is a schematic diagram of the DEA-Unit module training / inference phase structure provided in Example 1. Figure 4 This is a schematic diagram of the DEA-Arch module structure provided in Example 1. Figure 5 This is a schematic diagram of the UBEH detection head structure provided in Embodiment 1. Figure 6 This is a schematic diagram of the data annotation interface and output effect provided in Example 1. Figure 7 This is a visualization comparison of key point detection between DU-YOLO and YOLO13-Pose in a complex industrial scenario, as provided in Example 1. Detailed Implementation

[0017] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0018] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and system for precise measurement of object dimensions based on the DU-YOLO architecture.

[0019] Example 1

[0020] This embodiment provides a method for precise measurement of object dimensions based on the DU-YOLO architecture, such as... Figures 1-2 As shown, it includes: Step S1. Obtain the image dataset of the object, and perform annotation and augmentation processing on the obtained image dataset to obtain the augmented dataset; Step S2. Construct an improved DU-YOLO detection model, wherein the improved DU-YOLO detection model is based on the YOLO13-Pose network as the baseline model, and introduces the dynamic evolution aggregation architecture DEA-Arch structure in the backbone network to replace the original backbone network, and introduces the practical balanced efficiency detection head UBEH structure in the detection head to replace the original detection head. Step S3. Use the augmented dataset from step S1 to train the improved DU-YOLO keypoint detection model to obtain the trained keypoint detection model; Step S4. Input the image of the object to be tested into the trained key point detection model to perform key point detection and obtain key point detection and precise size measurement results.

[0021] The detection method in this embodiment can perform non-contact automatic measurement of the dimensions of precision components (such as chips, connectors, capacitors, etc.) on electronic manufacturing production lines, including length, width, height, and pin spacing, with a measurement accuracy requirement of sub-pixel level (better than 0.5 pixels).

[0022] In step S1, an image dataset of the object is obtained, and the obtained image dataset is labeled and augmented to obtain an augmented dataset.

[0023] To achieve robustness of the model in complex industrial environments (variable lighting, partial occlusion, cluttered backgrounds), a high-quality dataset with depth information needs to be constructed, specifically: S11. Acquire an image containing the object and its corresponding depth map.

[0024] An Orbbec Astra S RGB-D camera was installed in an actual production line environment. This camera can acquire images of various objects in the actual factory production line environment and output raw RGB color images with a resolution of 1280×720 and corresponding depth maps. The value of each pixel in the depth map is the actual distance (in millimeters) from that point to the camera. Images were taken from multiple angles and under different lighting conditions for components of different sizes and categories.

[0025] S12. Perform the first data enhancement processing on the acquired image.

[0026] The acquired raw images and their depth maps undergo various data enhancement operations, including random rotation, horizontal / vertical flipping, brightness and contrast adjustment, and random occlusion, to expand data diversity and simulate real disturbances. Specific operations include: The image is rotated at a randomly selected angle within the range of [-15°, +15°], and the depth map is rotated synchronously.

[0027] The image is flipped with a probability of 0.5 to simulate different orientations of the components.

[0028] The brightness and contrast of the RGB image are randomly adjusted within the range of [0.8, 1.2] times the original value to simulate changes in lighting.

[0029] Randomly place 2 to 4 gray rectangular blocks in the image, covering no more than 20% of the total image area, to simulate debris obstruction that may occur on the production line.

[0030] S13. Add bounding boxes and key point annotations to the first enhanced image.

[0031] The open-source Labelme annotation tool was used to annotate the bounding boxes and key points of each enhanced image.

[0032] Bounding box annotation involves drawing a rectangular box tightly enclosing the outline of each component under test in the image and labeling its category. Key point annotation involves marking preset feature key points on each instance according to the dimensional measurement requirements. For example, for a rectangular capacitor, its four corner points (a total of 4 key points) are marked for calculating its length and width; for a QFP packaged chip, the tips of its outer pins are marked (e.g., 16 key points) for calculating the pin spacing. In this embodiment, the total number of key points is set to vary from 4 to 16 depending on the complexity of the object. Ultimately, approximately 10,000 instances are annotated, generating more than 60,000 key point coordinates.

[0033] S14. Perform a second enhancement process on the labeled images to obtain the enhanced dataset.

[0034] To further improve the model's generalization ability to image noise and color distortion, the labeled images are enhanced again using Gaussian blur, Gaussian noise addition, color dithering, and random erasure. This round of enhancement does not change the geometric coordinates of the labels. Specific operations include: The image was slightly blurred using a Gaussian kernel with a standard deviation of [0, 1.5].

[0035] Add Gaussian noise with a mean of 0 and a standard deviation in the range [0, 0.02*255] to the RGB image.

[0036] In the HSV color space, the hue (H), saturation (S), and lightness (V) channels are subjected to small random perturbations (±5%).

[0037] Randomly erase a small rectangular region (area ≤ 10%) from the image and fill it with the image mean.

[0038] The final result was an augmented dataset consisting of approximately 3024 images and their corresponding bounding box labels, keypoint labels, and depth maps.

[0039] In step S2, an improved DU-YOLO detection model is constructed. The improved DU-YOLO detection model uses the YOLO13-Pose network as the baseline model. The original backbone network is replaced by a dynamically evolving aggregation architecture DEA-Arch structure, and the original detection head is replaced by a practical balanced efficiency detection head UBEH structure.

[0040] This embodiment addresses the challenges of numerous small targets and limited computational resources in industrial scenarios by reconstructing the backbone network and detection head of the baseline model YOLO13-Pose. Specifically: S21. The publicly available YOLO13-Pose network is used as the baseline model. Its original backbone network is CSPDarknet, and the detection head is a convolutional layer that couples classification, regression and key point prediction tasks.

[0041] S22. Introduce a dynamic evolutionary aggregation architecture (DEA-Arch) to the backbone network to replace the original backbone network CSPDarknet; the dynamic evolutionary aggregation architecture (DEA-Arch) includes a dynamic evolutionary unit (DEA-Unit) and an Arch single-aggregation mechanism, such as... Figure 3 , Figure 4 As shown.

[0042] The dynamic evolution unit consists of a dimensionality reduction convolutional layer, a multi-branch feature extraction layer, and a channel rearrangement layer connected in sequence.

[0043] Training phase: The feature map X (size H×W×C) from the previous layer is received, and the dimensionality of the input feature is reduced to C / 2 by a 1×1 convolutional layer. Subsequently, the feature is fed into a multi-branch feature extraction layer.

[0044] The multi-branch feature extraction layer retains at least 2×2 convolutional branches, 1×1 convolutional branches, and residual connection branches during the training phase, and merges them into a single 2×2 convolutional layer through structural reparameterization during the inference phase.

[0045] The output feature maps of the three branches are concatenated along the channel dimension and then exchanged through a channel shuffle layer. This design greatly enhances the diversity and expressive power of the features during training.

[0046] Inference phase: Through structural reparameterization, the parallel 2×2 convolutional branches, 1×1 convolutional branches, and residual connection branches are mathematically equivalently merged into a single 2×2 convolutional layer. This significantly reduces the number of model parameters and the computational cost during inference, while significantly improving the ability to represent small target features.

[0047] The single-step aggregation mechanism is connected after the dynamic evolution unit and is configured to receive the output of the channel rearrangement layer and perform a one-time multi-scale feature aggregation, specifically: The features output by the dynamically evolving unit are fed into an Arch single-pass aggregation module. This module employs a one-time aggregation strategy, capturing multi-scale contextual information by using dilated convolutions or pooling operations with different dilation rates in parallel, and fusing the output feature map X1 in one go. Compared to the complex path of multiple downsampling and upsampling in the original network, this mechanism is more efficient.

[0048] S23. To address the issues of insufficient accuracy in key point positioning and conflicting features among multiple tasks, a practical balanced efficiency detection head (UBEH) structure is introduced to replace the original coupled detection head.

[0049] The practical balanced efficiency detection head UBEH structure includes a pyramid-shaped multi-scale feature fusion module, a task decoupling module, an adaptive deformation module, and a feature adaptation module.

[0050] The pyramid-shaped multi-scale feature fusion module connects its input to the output of the backbone network to receive and fuse feature maps from different scales of the backbone network. Its multiple outputs connect to the inputs of the task decoupling layers to receive three scale feature layers from the backbone network's DEA-Arch output: C3 (large scale, rich in detail), C4 (medium scale), and C5 (small scale, strong semantic information). First, the C3, C4, and C5 feature maps are upsampled to the same size using a weighted concatenation unit and then weighted and concatenated according to a channel ratio of 4:3:2. Subsequently, the concatenated features pass through a spatial pyramid fast fusion module (SPFF). This module efficiently aggregates global contextual information through parallel pooling and convolution operations, outputting enhanced multi-scale features (X2), which are then fed into the three decoupling branches.

[0051] The task decoupling module, built on the Transformer decoder architecture, takes X2 as input and processes it through three completely independent feedforward network branches (implemented based on RT-DETR Decoder). These three completely independent feedforward network branches include a classification branch, a bounding box regression branch, and a key point localization branch. These three branches share features but do not share parameters during training, completely decoupling the learning objectives of different tasks and avoiding feature conflicts.

[0052] The classification branch outputs the class confidence score for each detection box; the bounding box regression branch outputs the center point coordinates and width and height offsets of the detection box; and the key point localization branch outputs the normalized coordinate offset of each preset key point relative to the center point of its corresponding detection box.

[0053] The adaptive deformation module has a first input terminal and a second input terminal, which are respectively connected to the internal feature nodes in the task decoupling module corresponding to the bounding box regression information and the key point localization information.

[0054] The feature adaptation module, connected to the adaptive deformation module, is used to provide the adaptive deformation module with learnable offset parameters and modulation mask parameters.

[0055] like Figure 5 As shown, this embodiment introduces an Adaptive Deformation Module (ADM) to achieve dynamic routing in the bounding box regression and keypoint localization branches. The core of this module is a deformable convolutional layer whose feature sampling positions are not fixed grids, but rather offsets predicted in real-time by the Feature Adaptation Module (FAM). Simultaneously, the Feature Adaptation Module adaptively generates learnable offset parameters and modulation mask parameters for each sampling point, and uses a transparent glass-like channel selector to globally adjust channel weights to weight the feature contribution of that point. Both parameters are learned from the current features through a small network, enabling the model to dynamically focus its attention on feature regions more important for bounding box and keypoint localization, thus achieving dynamic feature routing.

[0056] The final model constructed is the DU-YOLO detection model, whose parameter count was reduced from the original 2.82M to approximately 2.55M.

[0057] In step S3, the enhanced dataset from step S1 is used to train the improved DU-YOLO keypoint detection model, resulting in a trained keypoint detection model.

[0058] S31. The augmented dataset obtained in step S1 is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0059] S32. Training was performed on a server equipped with an NVIDIA RTX 4070 Super GPU, with the input image resolution fixed at 640×640. A stochastic gradient descent optimizer with momentum was used, with an initial learning rate of 0.01 and a cosine annealing strategy; the batch size was set to 16; and the total number of training epochs was 300–500.

[0060] S33. In each epoch, the model performs forward and backward propagation on the training set to update the weights. Every few epochs, an evaluation is performed on the validation set to calculate the mean accuracy (mAP) of keypoint detection. During training, the model weights with the highest mAP on the validation set are saved as the optimal weights for the DU-YOLO keypoint detection model.

[0061] The best DU-YOLO keypoint detection model trained was quantitatively compared with the original YOLO13-Pose model on the same test set, as shown in Table 1.

[0062] Table 1 Comparison of different algorithms

[0063] As shown in Table 1, the best DU-YOLO keypoint detection model reduces the number of parameters by 9.58% and the computational cost (GFLOPs) by 6.49%, while improving the average accuracy mAP50 of keypoint detection by 4.0 percentage points and the mAP50-95 by 3.1 percentage points. Its overall performance is significantly better than the original YOLO13-Pose.

[0064] In step S4, the image of the object to be tested is input into the trained key point detection model to perform key point detection, and the key point detection and precise size measurement results are obtained.

[0065] S41. Input an untrained image of the component under test from the production line into the trained DU-YOLO keypoint detection model. After inference, the model outputs a series of keypoint detection results. Each keypoint detection result is structured data, including: bounding box coordinates [x_center, y_center, width, height], class confidence (e.g., 0.98), and a list of N keypoints, each in the format [u, v, visibility], where u and v are the pixel coordinates of the keypoint in the input image, and visibility is the visibility confidence. The annotation interface and final output effect are shown below. Figure 6 As shown.

[0066] S42. Calculate the precise size measurement results based on the pixel coordinates of the key points, camera intrinsic parameters, and depth map information.

[0067] Based on the pixel coordinates (u, v) of the keypoint, find the corresponding depth value d (in meters) in the depth map acquired by S1 and strictly registered with the RGB image. Combine this with the camera's factory-calibrated intrinsic parameter matrix K (including focal length f). x , f y and principal point c x , c y Using the inverse transformation of the perspective projection model, calculate the coordinates (X, Y, Z) of the point in the camera's 3D space: Z=d; X=(u - c x ) * Z / f x ; Y=(v - c y ) * Z / f y ; Perform this calculation for each pair of key points whose dimensions need to be calculated, to obtain their coordinates (X1, Y1, Z1) and (X2, Y2, Z2) in 3D space.

[0068] S43. Calculate the Euclidean distance between the two points mentioned above as the actual physical dimension: D=sqrt((X1-X2)²+ (Y1-Y2)²+ (Z1-Z2)²).

[0069] Select the three-dimensional coordinates of three key points A, B, and C, calculate vectors BA and BC, and then calculate the included angle ∠ABC using the dot product formula.

[0070] By following the above steps, the actual physical dimensions of the object to be measured, such as length, width, height, hole spacing, diameter, or angle, can be obtained, thus enabling automatic measurement of three-dimensional physical dimensions from two-dimensional images to sub-pixel level precision.

[0071] Figure 7 The image shows a visual comparison between the best DU-YOLO keypoint detection model and the original YOLO13-Pose keypoint detection model. Figure 7 The two images at the top and middle show the original YOLO13-Pose keypoint detection results. Figure 7 The two images at the bottom center are visualizations of the DU-YOLO keypoint detection model. Under small target conditions, the keypoint localization is significantly more accurate and stable, and the size measurement error is greatly reduced. It fully meets the sub-pixel accuracy measurement requirements of the production line, verifying that the method has higher accuracy, better robustness in complex industrial scenarios, and practical application value, thus ensuring a smaller final size measurement error.

[0072] Example 2

[0073] This embodiment provides a precision measurement system for object dimensions based on the DU-YOLO architecture, including: The acquisition module is used to acquire image datasets of objects and perform annotation and augmentation processing on the acquired image datasets to obtain augmented datasets; The building module is used to build an improved DU-YOLO detection model. The improved DU-YOLO detection model is based on the YOLO13-Pose network. The original backbone network is replaced by a dynamically evolving aggregation architecture DEA-Arch structure, and the original detection head is replaced by a practical balanced efficiency detection head UBEH structure. The training module is used to train the improved DU-YOLO keypoint detection model using an augmented dataset to obtain a trained keypoint detection model. The detection module is used to input the image of the object to be measured into the trained key point detection model to perform key point detection and obtain key point detection and precise size measurement results.

[0074] It should be noted that the object size precision measurement system based on the DU-YOLO architecture provided in this embodiment is similar to that in Embodiment 1, and will not be described in detail here.

[0075] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A method for precise measurement of object dimensions based on the DU-YOLO architecture, characterized in that, include: Step S1. Obtain the image dataset of the object, and perform annotation and augmentation processing on the obtained image dataset to obtain the augmented dataset; Step S2. Construct an improved DU-YOLO detection model, wherein the improved DU-YOLO detection model is based on the YOLO13-Pose network as the baseline model, and introduces the dynamic evolution aggregation architecture DEA-Arch structure in the backbone network to replace the original backbone network, and introduces the practical balanced efficiency detection head UBEH structure in the detection head to replace the original detection head. Step S3. Use the augmented dataset from step S1 to train the improved DU-YOLO keypoint detection model to obtain the trained keypoint detection model; Step S4. Input the image of the object to be tested into the trained key point detection model to perform key point detection and obtain key point detection and precise size measurement results.

2. The method for precise measurement of object dimensions based on the DU-YOLO architecture according to claim 1, characterized in that, The DEA-Arch structure of the dynamic evolution aggregation architecture in step S2 includes: The dynamic evolution unit is used to reduce the dimensionality of the input features during the training phase. It extracts features by combining a multi-branch structure with channel rearrangement and merges them into a single-path convolution during the inference phase through structural reparameterization. The single-step aggregation mechanism is used to aggregate features output by dynamic evolution units to achieve multi-scale feature fusion and output.

3. The method for precise measurement of object dimensions based on the DU-YOLO architecture according to claim 2, characterized in that, The multi-branch structure in the dynamic evolution unit includes at least a 2×2 convolution branch, a 1×1 convolution branch, and a residual connection branch.

4. The method for precise measurement of object dimensions based on the DU-YOLO architecture according to claim 1, characterized in that, The UBEH structure of the practical balance efficiency detection head in step S2 includes: The task decoupling module is built on the Transformer decoder architecture and has separate first output interface, second output interface and third output interface, which are used to output target classification information, bounding box regression information and key point localization information, respectively. The adaptive deformation module has a first input end and a second input end, which are respectively connected to the internal feature nodes in the task decoupling module that correspond to the bounding box regression information and the key point localization information. The feature adaptation module, connected to the adaptive deformation module, is used to provide the adaptive deformation module with learnable offset parameters and modulation mask parameters.

5. The method for precise measurement of object dimensions based on the DU-YOLO architecture according to claim 4, characterized in that, The UBEH practical balance efficiency detection head structure also includes: The pyramid-shaped multi-scale feature fusion module has its input end connected to the output end of the backbone network. It is used to receive and fuse feature maps from different scales of the backbone network. Its multiple output ends are respectively connected to the task input ends of the task decoupling layer. It is used to perform weighted concatenation and context information enhancement processing on the feature layers of different scales output by the backbone network, and input the processed features to the task decoupling layer.

6. The method for precise measurement of object dimensions based on the DU-YOLO architecture according to claim 1, characterized in that, Step S1 specifically includes: Images containing objects and their corresponding depth maps are acquired. The acquired images undergo a first data augmentation process. Bounding boxes and keypoint annotations are added to the first augmented images. The annotated images undergo a second augmentation process to obtain the augmented dataset.

7. The method for precise measurement of object dimensions based on the DU-YOLO architecture according to claim 6, characterized in that, The first round of enhancement processing includes at least one of rotation, flipping, brightness adjustment, and random occlusion; the second enhancement processing includes at least one of blurring, adding noise, color jittering, and random erasure.

8. The method for precise measurement of object dimensions based on the DU-YOLO architecture according to claim 5, characterized in that, The key point detection results in step S4 include at least the bounding box coordinates of the object, the category confidence score, and the pixel coordinates of the key points in the image of the object to be tested.

9. The method for precise measurement of object dimensions based on the DU-YOLO architecture according to claim 8, characterized in that, The precise dimensional measurement results in step S4 are calculated based on the pixel coordinates of key points, camera intrinsic parameters, and depth map information.

10. A measurement system based on the DU-YOLO architecture-based method for precise measurement of object dimensions as described in any one of claims 1-9, characterized in that, include: The acquisition module is used to acquire image datasets of objects and perform annotation and augmentation processing on the acquired image datasets to obtain augmented datasets; The building module is used to build an improved DU-YOLO detection model. The improved DU-YOLO detection model is based on the YOLO13-Pose network. The original backbone network is replaced by a dynamically evolving aggregation architecture DEA-Arch structure, and the original detection head is replaced by a practical balanced efficiency detection head UBEH structure. The training module is used to train the improved DU-YOLO keypoint detection model using an augmented dataset to obtain a trained keypoint detection model. The detection module is used to input the image of the object to be measured into the trained key point detection model to perform key point detection and obtain key point detection and precise size measurement results.

Images