Pulse signal recognition method based on improved YOLOv5

Abstract

The application discloses an oscilloscope pulse signal recognition method based on improved YOLOv5, relates to the field of computer vision, and is particularly used for improving the recognition precision and efficiency of an oscilloscope pulse signal.The application optimizes YOLOv5 by structural change: a reduced ResNet50 is used to replace the original structure as a backbone network, details of feature maps are effectively reserved, and the detection accuracy of a pulse signal is improved; a bidirectional feature pyramid network (BiFPN) is introduced into a neck network, information fusion between feature maps of different scales is enhanced, and the transmission efficiency of local detail information is improved; meanwhile, the resolution and quantity of feature maps are adjusted, the use of high-resolution feature maps is increased, the proportion of low-resolution feature maps is reduced, and the sensitivity of the model to the pulse signal is further improved.In addition, the application also proposes an efficient data enhancement method, the data set is expanded by inserting the scaled and linearly transformed small data set into the background, the data acquisition cost is reduced, the diversity and scale of the data set are improved, and the generalization ability of the model is enhanced.

Description

Technical Field

[0001] This invention relates to the field of computer vision and signal processing technology, specifically a pulse signal recognition method based on an improved YOLOv5 for improving the accuracy and efficiency of pulse signal recognition in oscilloscopes. Background Technology

[0002] Object detection is a crucial task aimed at identifying and locating target objects in images or videos. YOLOv5 (You Only Look Once version 5), as an advanced single-stage object detection model, has gained widespread recognition in the field due to its superior performance and speed. YOLOv5's architecture is ingenious and flexible; its modular design allows developers to easily split, adjust, and reconstruct the model according to specific needs. This high degree of flexibility enables YOLOv5 to adapt well to various complex application scenarios, resulting in outstanding performance in real-world applications.

[0003] However, while YOLOv5 performs excellently in general object detection tasks, its performance may not be ideal in specific domains, such as pulse signal detection. This is mainly due to the following reasons:

[0004] As a general-purpose object detector, YOLOv5 is not specifically optimized for impulse signal detection. Impulse signals typically have unique characteristics and patterns, requiring the model to capture these subtle differences. Current optimizations of YOLOv5 mostly focus on adjusting specific parameters or enhancing the training set. While these methods can improve performance to some extent, they do not fundamentally address the model's shortcomings in impulse signal detection.

[0005] Furthermore, datasets play a crucial role in model training. A high-quality, large-scale dataset can significantly improve model performance and accuracy. However, in practical applications, acquiring and creating datasets faces numerous challenges. On the one hand, dataset annotation is time-consuming and labor-intensive, requiring specialized knowledge and skills; on the other hand, in certain sensitive fields (such as medical imaging), data acquisition and annotation are extremely costly and may even be subject to legal and privacy policy restrictions. These factors all contribute to the difficulty in obtaining sufficiently large datasets to train models, thus limiting further improvements in model performance to some extent.

[0006] In summary, although YOLOv5 has achieved remarkable success in general object detection tasks, it still has shortcomings in specific areas such as pulse signal detection and in dataset acquisition and creation.

[0007] Therefore, there is a need for a YOLOv5 model optimized for pulse signal detection and a more efficient method for creating datasets to address the problems existing in the current technology. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes an optimization method for the YOLOv5 model, aiming to improve the model's performance in pulse signal recognition through structural changes. Simultaneously, it proposes an efficient data augmentation method to reduce data acquisition costs, improve the quality and scale of the dataset, thereby further enhancing the model's performance.

[0009] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0010] A pulse signal recognition method based on an improved YOLOv5, characterized by the following steps:

[0011] S1. Select an image of an oscilloscope interface as the background;

[0012] S2. Scale the images in the labeled dataset;

[0013] S3. Perform certain transformations on the scaled image, such as translation and rotation, and insert it into the background image to expand the dataset;

[0014] S4 divides the expanded dataset into a training set, a test set, and a validation set;

[0015] S5. Establish an improved YOLOv5 network model, which consists of a backbone network, a neck network, and an output terminal;

[0016] S6. Input the images from the training set into the backbone network of the improved YOLOv5 network model, and obtain the feature output. Figure 1 ;

[0017] S7. Features Figure 1 The input is fed into the neck network of the improved YOLOv5 network model, and the output is the feature. Figure 2 ;

[0018] S8. Features Figure 2 The input is fed into the output of the improved YOLOv5 network model, and the output is the detection box of the pulse signal in the oscilloscope image.

[0019] The improved YOLOv5 model is special in that:

[0020] The YOLOv5 backbone network was improved by replacing the original structure with ResNet50 and reducing the size of the DenseNet component. ResNet50, through its dense connections, more effectively preserves the detailed information of the feature maps, which is crucial for impulse signal detection. Experimental results also show that reducing the size of DenseNet can significantly improve the accuracy of impulse signal detection while maintaining low computation time.

[0021] The neck structure of YOLOv5 was adjusted by employing a Bidirectional Feature Pyramid Network (BiFPN) to enhance information fusion between feature maps of different scales, improving the efficiency of local detail transmission and reducing information loss during convolution. This improvement helps the model better capture detailed features, thereby enhancing detection performance.

[0022] The resolution and number of feature maps were adjusted. By increasing the use of high-resolution feature maps, the pixel proportion of small objects in the feature maps was increased, thus making their impact on the model more significant. At the same time, the use of low-resolution feature maps was reduced, lowering the computational burden on the model.

[0023] In addition, a high-efficiency data augmentation system is provided for expanding datasets, characterized by comprising:

[0024] Background selection module: Filter background images containing various oscilloscope interface types based on the application scenario;

[0025] Signal conversion module: Performs operations on pulse signal images, including scaling to a preset resolution, panning within ±20% of the image size, rotating within ±15°, and adjusting brightness within ±20%.

[0026] Synthesis module: Inserts the transformed signal into the non-overlapping area of ​​the background image, avoiding the coordinate axis area, and automatically generates annotation box information through a coordinate mapping algorithm.

[0027] Compared with the prior art, the present invention has the following effects:

[0028] 1. This invention replaces the original CSPDarknet backbone network of YOLOv5 with ResNet50. By reducing the number of DenseBlock layers and channels, it reduces computational complexity while retaining the advantages of dense connections. Furthermore, it combines a Bidirectional Feature Pyramid Network (BiFPN) to connect multi-scale feature maps via bidirectional paths from top to bottom and bottom to top, achieving efficient interaction between high-frequency details and global semantic information, effectively improving the model's feature extraction capability.

[0029] 2. By adjusting the depth and width of the model, the resolution of the feature map is improved, thereby enhancing the model's sensitivity to pulse signals and its detection accuracy.

[0030] 3. This invention effectively expands the dataset by scaling and linearly transforming a pre-labeled small dataset and inserting it into a background image. It simplifies the labeling process, reduces manual annotation costs, and improves the model's generalization ability. Attached Figure Description

[0031] Figure 1 This is a schematic diagram illustrating the structural principle of an improved YOLOv5 model for pulse signal recognition according to the present invention.

[0032] Figure 2 This is a tagged pulse signal image from an embodiment of the present invention;

[0033] Figure 3 The image shows a pulse signal image and its label after data enhancement, as described in an embodiment of the present invention. Detailed Implementation

[0034] The present invention will be further described below with reference to implementation examples and accompanying drawings, but this should not be construed as limiting the scope of protection of the present invention.

[0035] A pulse signal recognition method based on an improved YOLOv5 includes the following steps:

[0036] S1. Select an oscilloscope interface image as the background;

[0037] S2. Scale the images in the labeled dataset to a fixed size at the target resolution while maintaining the aspect ratio;

[0038] S3. Perform a linear transformation on the scaled image and insert the transformed pulse signal into the background image to expand the dataset;

[0039] S4 divides the expanded dataset into a training set, a test set, and a validation set;

[0040] S5. Construct an improved YOLOv5 network model, which consists of a backbone network, a neck network, and an output network. The backbone network uses a ResNet50 structure and reduces the size of DenseNet, preserving detailed information of the feature maps through dense connections. The neck network introduces a Bidirectional Feature Pyramid Network (BiFPN) to enhance information fusion between feature maps of different scales. The output network adjusts the resolution and number of feature maps, increasing the proportion of high-resolution feature maps and decreasing the proportion of low-resolution feature maps.

[0041] S6. Input the images from the training set into the backbone network of the improved YOLOv5 network model, and obtain the feature output. Figure 1 ;

[0042] S7. Features Figure 1 The input is fed into the neck network of the improved YOLOv5 network model, and the output is the feature. Figure 2 ;

[0043] S8. Features Figure 2 The input is fed into the output of the improved YOLOv5 network model, and the output is the detection box of the pulse signal in the oscilloscope image.

[0044] Steps S5-8 include an improved YOLOv5 model for pulse signal recognition, as invented, such as... Figure 1 As shown, the specific steps include:

[0045] a. Adjust the input image to 640*640 pixels and input it into the backbone network;

[0046] b. Input the image into the first feature extraction block, which includes the Focus module, Conv module and DenseBlock module, to obtain the feature map Fm1 (160*160 pixels);

[0047] c. Input feature map Fm1 into the second feature extraction block, which includes the Conv module and the DenseBlock module, to obtain feature map Fm2 (80*80 pixels);

[0048] d. Input feature map Fm2 into the third feature extraction block, which includes the Conv module and the DenseBlock module, to obtain feature map Fm3 (40*40 pixels);

[0049] e. Input feature map Fm3 into the fourth feature extraction block, which includes the Conv module and the SPP module, to obtain feature map Fm4 (20*20 pixels);

[0050] f. Input the feature map Fm4 into the upsampling block in the neck network, which includes the Conv module, Upsample module and CSP module, to obtain the feature map Fm5 (40*40 pixels);

[0051] g. Input feature maps Fm5 and Fm3 into the first feature fusion block, which includes the Conv module, Upsample module, Concat module and CSP module, and fuse the features to obtain feature map Fm6 (80*80 pixels);

[0052] h. Input feature maps Fm6 and Fm2 into the second feature fusion block, which includes the Conv module, Upsample module, Concat module and CSP module, and fuse the features to obtain feature map Fm7 (160*160 pixels);

[0053] i. Input feature maps Fm7 and Fm1 into the third feature fusion block, which includes the Conv module, Upsample module, Concat module and CSP module, and fuse the features to obtain feature map Fm8 (160*160 pixels);

[0054] j. Input feature maps Fm8 and Fm7 into the fourth feature fusion block, which includes the Conv module, Concat module and CSP module, and fuse the features to obtain feature map Fm8 (80*80 pixels);

[0055] k. Input feature maps Fm9 and Fm6 into the fifth feature fusion block, which includes the Conv module, Concat module and CSP module, and fuse the features to obtain feature map Fm10 (40*40 pixels);

[0056] L. Input the feature maps Fm7, Fm8 and Fm9 into different Detect modules in the head network to predict the target, and output the detection box of the pulse signal.

[0057] Steps S1-3 comprise an efficient data augmentation method of the invention, the process of which is as follows: Figure 2 As shown, the specific implementation steps are as follows:

[0058] a. Obtain n images of pulse signals;

[0059] b. Extract and label the signal portion to create a dataset, such as... Figure 3 Its dimensions are 72*48 pixels;

[0060] c. Select an oscilloscope interface image as the background. Here, the image size is 800*480 pixels.

[0061] d. Scale the images in the dataset to a size smaller than 267*160 pixels while maintaining the aspect ratio. In this example, the image size already meets the requirements, so no scaling is required.

[0062] e. Insert the images in the dataset into the background at positions (50,50), (50,167), (50,283), (333,50), (333,167), (333,283), (617,50), (617,167), and (617,283), respectively.

[0063] f. Add (50,50), (50,167), (50,283), (333,50), (333,167), (333,283), (617,50), (617,167), and (617,283) to the bounding box positions in the labels of the dataset to obtain a new dataset.

[0064] Experiments show that this invention, through deep optimization of the YOLOv5 model structure, significantly enhances the model's feature extraction and detection capabilities, enabling more accurate capture and identification of pulse signals in oscilloscopes. Simultaneously, innovative data augmentation methods effectively expand the dataset, simplifying the tedious labeling process and enhancing sample diversity and richness, further improving the model's generalization ability and robustness, and effectively increasing the accuracy and efficiency of oscilloscope pulse signal recognition.

Claims

1. A pulse signal recognition method based on an improved YOLOv5, characterized in that, The method includes the following steps: S1. Select an oscilloscope interface image as the background; S2. Scale the images in the labeled dataset to a fixed size at the target resolution while maintaining the aspect ratio; S3. Perform a linear transformation on the scaled image and insert the transformed pulse signal into the background image to expand the dataset; S4 divides the expanded dataset into a training set, a test set, and a validation set; S5. Construct an improved YOLOv5 network model, which consists of a backbone network, a neck network, and an output network. The backbone network uses a ResNet50 structure and reduces the size of DenseNet, preserving detailed information of the feature maps through dense connections. The neck network introduces a Bidirectional Feature Pyramid Network (BiFPN) to enhance information fusion between feature maps of different scales. The output network adjusts the resolution and number of feature maps, increasing the proportion of high-resolution feature maps and decreasing the proportion of low-resolution feature maps. S6. Input the images in the training set into the backbone network of the improved YOLOv5 network model, and output feature map 1; S7. Input feature map 1 into the neck network of the improved YOLOv5 network model and output feature map 2; S8. Input feature map 2 into the output of the improved YOLOv5 network model to obtain the detection box of the pulse signal in the oscilloscope image.

2. The pulse signal recognition method based on improved YOLOv5 according to claim 1, characterized in that, In the S3 data augmentation step, the pulse signal insertion position avoids the coordinate axis area of ​​the oscilloscope interface, and a label box is automatically generated through a coordinate mapping algorithm. The coordinates of the label box are dynamically adjusted according to the offset of the insertion position.

3. The pulse signal recognition method based on improved YOLOv5 according to claim 1, characterized in that, The S3 transformation operation includes at least one of translation, rotation, scaling, and flipping.

4. The pulse signal recognition method based on improved YOLOv5 according to claim 1, characterized in that, The selection of the background image should fully consider the actual application scenario and the characteristics of the target object in order to simulate the situation in the real environment and improve the generalization ability and adaptability of the model.

5. A data augmentation system for expanding an oscilloscope pulse signal dataset, characterized in that, include: Background selection module: Filter background images containing various oscilloscope interface types based on the application scenario. Signal transformation module: Performs scaling, translation, and rotation operations on pulse signal images; Synthesis module: Inserts the transformed signal into the non-overlapping area of ​​the background image, avoiding the coordinate axis area, and automatically generates annotation box information through a coordinate mapping algorithm.

6. The data augmentation system according to claim 5, characterized in that, The coordinate mapping algorithm dynamically generates the coordinates of the upper left and lower right corners of the annotation box by calculating the center point coordinates and size of the transformed pulse signal and combining them with the insertion position of the background image.

Images