A submarine cable target detection method based on a YOLO-CABLE model

Abstract

The application discloses a submarine cable target detection method based on a YOLO-CABLE model. The YOLO-CABLE model is obtained by improving a basic YOLO-V3 network model. In the feature extraction network, the output of the first residual block is respectively subjected to convolution operation and spliced with the output of the third and fourth residual blocks, and then input into the next residual block, so that the position information of the shallow layer is transmitted to the high layer. In the prediction network, the large-scale branch is removed, the network structure is lightened, and the detection speed is improved. In the feature fusion network, the top-down and bottom-up bidirectional paths are adopted to fuse the high-level features. Finally, the submarine image is detected by using the trained YOLO-CABLE model. Compared with the prior art, the average precision is improved by 4.2%, and the average detection speed is shortened by 1.616 seconds.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology and relates to deep learning based on image detection models, specifically to a method for detecting submarine cable targets based on the YOLO-CABLE model. Background Technology

[0002] With the rapid growth of the world's population, the development of marine resources has received increasing attention. Submarine cables, as an indispensable part of marine resource development and island economic development, have received widespread attention. The seabed environment is extremely complex, and irreversible damage events such as seabed erosion and ship anchor damage can cause serious damage to submarine cables. To ensure the normal operation of submarine cables, regular inspections are necessary. With the continuous advancement of underwater technology and the increasing demands for inspection efficiency, submarine cable inspection methods have shifted from traditional divers to intelligent underwater robots. Machine vision technology plays a crucial role in underwater robot inspections. Currently, underwater robots using machine vision for submarine cable inspection has become routine.

[0003] During the inspection, the underwater robot is in motion, resulting in blurry images of the submarine cables captured by the camera. This leads to a lack of location and feature information for the cables in the images. Furthermore, the dim underwater lighting and the influence of the light propagation medium cause partial distortion in the image content, giving it a bluish-green tint. Existing technologies use common target detection algorithms such as YOLOv3, Faster R-CNN, and SSD to detect submarine cables in images. These algorithms train convolutional neural network models on the cable images to detect information such as the cable's location, number, and size. However, blurry images can make it difficult for algorithms to accurately identify the cable's location and size; color distortion can also affect the accuracy of algorithms that use color features for detection; and dim lighting reduces the camera's resolution and sensitivity, thus affecting detection performance. These inherent disadvantages negatively impact the algorithm's recognition results, reducing accuracy.

[0004] Improving image quality through noise reduction or color feature enhancement can increase recognition accuracy. Acquiring more precise submarine cable location information during image capture using methods such as laser scanning and acoustic detection also contributes to improved accuracy. However, these methods require additional resource consumption. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a submarine cable target detection method based on the YOLO-CABLE model. It improves the feature transmission channel of the feature extraction network and the feature fusion path of the fusion network, thereby enhancing the detection accuracy and speed of the model.

[0006] A method for detecting submarine cables based on the YOLO-CABLE model includes the following steps:

[0007] Step 1: Collect seabed image data, remove images without submarine cable targets, perform data augmentation on the remaining images to expand the number of images, and manually label the submarine cable targets in the images as corresponding labels to create a training dataset.

[0008] Step 2: Build a YOLO-V3 network model, and then improve upon it to construct a YOLO-CABLE model. The specific improvement steps are as follows:

[0009] s2.1 In the feature extraction network, the first residual block belongs to the lowest feature layer and has good target location information. As the convolution depth increases, the higher feature layers have rich semantic information but lack target location information. Therefore, skip connections are added to the feature extraction network of the YOLO-V3 network model to enhance the target location information in higher layers. Specifically, the output of the first residual block is convolved and then concatenated with the output of the third residual block as the input of the fourth residual block. Then, the output of the first residual block is convolved again and concatenated with the output of the fourth residual block as the input of the fifth residual block.

[0010] By using skip connections, low-level features containing target location information are passed into high-level feature layers, so that the output of the feature extraction network not only has rich feature semantic information, but also contains good target location information.

[0011] The prediction network of s2.2 and YOLO-V3 includes three branches, dividing the image into 13x13, 26x26, and 52x52 grids respectively for detecting large, medium, and small targets. Considering that the detected target, the submarine cable, is a linear object and usually occupies a large proportion of the image, the branch for detecting small targets would lead to network redundancy. Therefore, the 52x52 branch used for detecting small targets in the YOLO-V3 network model was removed, reducing unnecessary computation, lightweighting the network structure, and accelerating the detection speed.

[0012] s2.3 Considering the complexity of seabed images and the importance of target features and locations, a bidirectional path of top-down and bottom-up is adopted in the multi-scale feature fusion network. The outputs of the fourth and fifth residual layers are fused, thereby upsampling higher-level semantic information to the lower-level output network where feature information is insufficient, and downsampling lower-level location information to the higher-level output network where location information is insufficient. Specifically, the output of the fifth residual layer is convolved, then upsampled and concatenated with the output of the fourth residual layer. After convolution, a large-scale target detection result is obtained. Then, the large-scale detection result is convolved and downsampled, concatenated with the convolution result before upsampling, and then convolved again to obtain a small-scale target detection result.

[0013] Step 3: Use the training dataset created in Step 1 to train the YOLO-CABLE model constructed in Step 2. Input the seabed image into the trained YOLO-CABLE model and output whether the submarine cable exists in the image and the location of the submarine cable.

[0014] The present invention has the following beneficial effects:

[0015] Skip connections are added to the YOLO-V3 feature extraction network to enhance the location information of the detected submarine cable. A top-down sampling structure is added to the multi-scale feature fusion to reduce network computation and expand the receptive field of higher-level feature layers. Simultaneously, considering the characteristics of submarine cables in images, the prediction network is lightweighted to improve detection speed. The overall structure compensates for the lack of feature information caused by removing small target detection. In comparative experiments with other existing technologies, the improved YOLO-CABLE model achieves a 4.2% improvement in average detection accuracy and a 1.616-second reduction in average detection speed. Experimental data demonstrate that our proposed YOLO-CABLE model surpasses the performance of other existing network models in the field of machine vision for submarine cable recognition. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the YOLO-CABLE model structure;

[0017] Figure 2 This is a schematic diagram of the feature extraction network structure in the embodiment;

[0018] Figure 3 This is a schematic diagram of the prediction network structure in the embodiment;

[0019] Figure 4 This is a schematic diagram of the feature fusion network structure in the embodiment;

[0020] Figure 5The examples show the loss curves of the YOLO-CABLE model under different datasets.

[0021] Figure 6 The results are shown in the example. Detailed Implementation

[0022] The present invention will be further explained below with reference to the accompanying drawings;

[0023] A method for detecting submarine cables based on the YOLO-CABLE model includes the following steps:

[0024] Step 1: PVC pipes were laid in the test pool at Hangzhou Dianzi University (Dongyue Campus) to simulate submarine cables. A camera was mounted on the bottom of an underwater vehicle to simulate an underwater robot inspecting the cables. The underwater vehicle moved at different speeds and captured images of the PVC pipes. To simulate a realistic underwater shooting environment, some interference factors were added, including motion blur, partial loss, occlusion, and the absorption and scattering effects of water on light. A total of 3104 images were captured. After removing images without submarine cable targets, the remaining 2399 images underwent brightness transformation, color balancing, and rotation processing, expanding the number of submarine cable images to 4798 without substantially increasing the image data. Then, submarine cable targets in the images were selected using horizontal bounding boxes as labels for the corresponding images. Finally, the labeled images were converted to PASCALVOC format and randomly assigned to training, validation, and test sets. The training set included 2886 images, the validation set included 432 images, and the test set included 480 images.

[0025] Step 2: Build a YOLO-V3 network model, and then improve upon it to construct, as shown below. Figure 1 The specific improvement steps for the YOLO-CABLE model shown are as follows:

[0026] s2.1, such as Figure 2 As shown, in the feature extraction network, the output of the first residual block undergoes a convolution operation with a kernel size of 3x3 and a stride of 4, resulting in a feature map of size 52x52 with 64 channels. This feature map is then concatenated with the output of the third residual block as the input to the fourth residual block. Next, the output of the first residual block undergoes a convolution operation with a kernel size of 6x6 and a stride of 8, resulting in a feature map of size 26x26 with 64 channels. This feature map is then concatenated with the output of the fourth residual block as the input to the fifth residual block.

[0027] s2.2 Considering that the detected target, the submarine cable, is a linear object and usually occupies a large proportion of the image, small target detection branches would lead to network redundancy. Therefore, in the YOLO-CABLE model, only the 26x26 and 13x13 branches are retained, such as... Figure 3 As shown.

[0028] s2.3, Feature fusion network, such as Figure 4 As shown, the output of the fifth residual block is first subjected to five convolution operations to obtain the feature map to be fused. Then, the number of channels is reduced by a single convolution operation with a kernel size of 1x1, and the size is expanded to 26x26 by a single upsampling operation. After being fused with the output of the fourth residual block, it is subjected to five more convolutions to obtain the 26x26 scale target detection result. At the same time, the 26x26 scale target detection result is subjected to a single convolution operation with a kernel size of 1x1 to increase the number of channels. Then, the size is reduced to 13x13 by a max pooling operation. After being concatenated with the fused feature map, it is subjected to five more convolutions to obtain the 13x13 scale target detection result.

[0029] Step 3: Conduct training and testing experiments on a desktop computer. The hardware parameters are shown in Table 1.

[0030]

[0031] Table 1

[0032] The YOLO-CABLE model constructed in step 2 was trained. Considering the impact of CPU and memory, the batch size was set to 8, and the training was performed for 100 epochs. The first 50 epochs were used for frozen training with a learning rate of 0.001. The feature extraction network was frozen, meaning the network structure remained unchanged, and only fine-tuning was performed. The last 50 epochs were then used for unfrozen training with a learning rate of 0.0001. After training, the corresponding weight file was saved. The YOLO-CABLE model was then trained using both the original captured image and the data-augmented image. The model loss was as follows: Figure 5 As shown.

[0033] Input validation set data into the trained YOLO-CABLE model, and output the presence and location of submarine cables in the image. The detection results are as follows: Figure 6 As shown, the boxes represent identified target submarine cables, and the value in the upper left corner represents the confidence score of the corresponding prediction box. The detection results show that the YOLO-CABLE model has a high recognition rate for submarine cables, achieving an accuracy of over 97% for straight cables. Even with curved shapes or obstructions, it still achieves good recognition accuracy.

[0034] Comparative experiments were conducted with other commonly used target detection models in existing technologies, and the results are shown in Table 2:

[0035]

[0036]

[0037] Table 2

[0038] As can be seen, the average detection accuracy of the YOLO-CABLE model described in this application is improved by 4.2%, and the average detection speed is shortened by 1.616 seconds, proving that the YOLO-CABLE model can surpass the performance of other commonly used network models in the field of machine vision in submarine cable identification.

Claims

1. A method for detecting submarine cable targets based on the YOLO-CABLE model, used to detect the presence and location of submarine cable targets in seabed images, characterized in that: Specifically, the following steps are included: Step 1: Collect seabed image data, remove images without submarine cable targets, perform data augmentation on the remaining images, manually label the locations of submarine cable targets in the images as corresponding labels, and create a training dataset. Step 2: Build a YOLO-V3 network model, and then improve upon it to construct a YOLO-CABLE model. The specific improvement steps are as follows: s2.1 Perform a convolution operation on the output of the first residual block in the feature extraction network, and concatenate it with the output of the third residual block as the input of the fourth residual block; then perform a convolution operation on the output of the first residual block, and concatenate it with the output of the fourth residual block as the input of the fifth residual block. s2.2 Remove branches with a scale of 52x52 from the prediction network of the YOLO-V3 network model; s2.3 In the feature fusion network, the output of the fifth residual block is convolved, then upsampled and concatenated with the output of the fourth residual layer. After convolution, a large-scale target detection result is obtained. Then, the large-scale detection result is convolved and downsampled, concatenated with the convolution result before upsampling, and then convolved again to obtain a small-scale target detection result. Step 3: Use the training dataset created in Step 1 to train the YOLO-CABLE model constructed in Step 2. Input the seabed image into the trained YOLO-CABLE model and output whether the submarine cable exists in the image and the location of the submarine cable.

2. The submarine cable target detection method based on the YOLO-CABLE model as described in claim 1, characterized in that: The data augmentation methods include brightness transformation, color balancing, and rotation processing.

3. The submarine cable target detection method based on the YOLO-CABLE model as described in claim 1, characterized in that: In the feature extraction network, the output of the first residual block is subjected to a convolution operation with a kernel size of 3x3 and a stride of 4, resulting in a feature map with a size of 52x52 and 64 channels. This feature map is then concatenated with the output of the third residual block as the input of the fourth residual block. Next, the output of the first residual block is subjected to a convolution operation with a kernel size of 6x6 and a stride of 8, resulting in a feature map with a size of 26x26 and 64 channels. This feature map is then concatenated with the output of the fourth residual block as the input of the fifth residual block.

4. The submarine cable target detection method based on the YOLO-CABLE model as described in claim 1, characterized in that: In the feature fusion network, the output of the fifth residual block is first subjected to five convolution operations to obtain the feature map to be fused. Then, the number of channels is reduced by a convolution operation with a kernel size of 1x1, and the size is expanded to 26x26 by an upsampling operation. After being fused with the output of the fourth residual block, it is subjected to five more convolutions to obtain the 26x26 scale target detection result. At the same time, the 26x26 scale target detection result is subjected to a convolution operation with a kernel size of 1x1 to increase the number of channels. Then, the size is reduced to 13x13 by a max pooling operation. After being concatenated with the fused feature map, it is subjected to five more convolutions to obtain the 13x13 scale target detection result.

5. The submarine cable target detection method based on the YOLO-CABLE model as described in any one of claims 1 to 4, characterized in that: The model is set to a batch size of 8, trained for 100 epochs, and frozen during the first 50 epochs with a learning rate of 0.001 to freeze the feature extraction network. After the last 50 cycles, the training was thawed again, with a learning rate of 0.0001.

6. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1 to 4.

Images