Forward-looking sonar image target detection method based on efficient pseudo tag

Abstract

The invention relates to a foresight sonar image target detection method based on an efficient pseudo tag, and the method comprises the steps: taking a Gaussian mixture model obtained by fitting an original sonar image as input, and providing the priori of sonar image distribution for a depth image priori network based on the Gaussian mixture model, thereby generating a high-quality de-noised sonar image. Meanwhile, a detection-friendly image quality evaluation network is constructed, and the best de-noised sonar image is selected from a group of de-noised sonar images generated by the depth image prior network as a pseudo tag to assist the backbone network in detection. A reconstruction module is used as a medium, and the whole detection network is optimized by a pseudo tag, so that a backbone network can extract features submerged by noise in an original input image. Compared with the prior art, the method has the advantages of high detection performance, high denoising capability, wide application range and the like.

Description

Technical Field

[0001] The present invention relates to the field of underwater image detection, and in particular to a forward-looking sonar image target detection method based on efficient pseudo-labels. Background Art

[0002] The sonar imaging system can truly reflect the underwater environment without being restricted by underwater lighting conditions, and plays an important role in the identification and detection of marine targets. Unlike optical images, acoustic images actively emit sound waves through sonar and map the pixel values ​​of the image according to the echo intensity. The pixel points corresponding to the areas with stronger echoes have stronger responses. However, the unique imaging characteristics of sonar equipment and the complex underwater environment make the captured images often full of noise, blurred targets, low contrast and other problems, which seriously affect the performance of target detection.

[0003] In recent years, with the development of deep learning, the performance of sonar image target detection algorithms based on convolutional neural networks (CNNs) has been greatly improved. For example, in solving the challenge of improving the accuracy of underwater sonar image detection using limited training data, some methods have achieved high accuracy through CNN networks, surpassing the performance of some template matching detection methods; some methods have designed a feature extraction network, using residual blocks to replace the backbone in the Mask RCNN network, reducing network parameters without affecting accuracy. In addition, in order to reduce network complexity and running time, single-stage networks such as YOLOv3, SSD, and RetinaNet are also used as basic detection networks. Some methods introduce attention mechanisms in YOLOv5 and propose real-time target detection algorithms TR-YOLOv5, which have achieved a certain degree of improvement; some methods design multi-scale convolution structures to enhance the network's perception of small target features and improve small target detection performance; some methods propose dual-path feature fusion networks for feature extraction, achieving robust and real-time sonar image detection. However, the above methods do not fully consider the characteristics of underwater sonar images such as grayscale, low contrast, and multiple noise interference, and are not suitable for underwater sonar image processing scenarios.

[0004] When various noises on sonar images are removed, objects can be detected more accurately. Among them, traditional methods for removing noise on sonar images usually require modeling specific noises based on a large number of parameters, and have poor versatility. Classic adaptive despeckling filters include LEE filters, KUAN filters, FROST filters, and SRAD filters. In addition, there is a two-dimensional finite impulse response (FIR) Wiener filter driven by an adaptive cuckoo search (ACS) algorithm that can remove Gaussian noise. However, the above traditional methods usually do not work well because they damage the details and textures of the original sonar images during the denoising process. Some underwater sonar image denoising uses deep learning methods. For example, the image despeckling convolutional neural network (ID-CNN), which assumes that the noise follows a gamma distribution and uses a component-by-component division residual layer to restore the speckle component. Another method proposes a self-supervised training strategy that simplifies the generation of training sets for removing speckle noise from acoustic images. However, these methods are based on the assumption that speckle noise mainly appears on sonar images, without considering the complex noise distribution in real sonar images. In addition, due to the lack of noise-free sonar images, supervised methods need to synthesize the dataset with optical images, which introduces domain shift and has a certain impact on the denoising performance. At the same time, these methods only model the main noise on sonar images (such as speckle noise or reverberation noise), which has a known simple distribution.

[0005] In addition, although visually, denoised sonar images appear smoother and more suitable for human viewing, denoising sonar images before feeding them into the detection model does not necessarily help improve the performance of the detection model. Because the purposes of visual denoising and detection denoising are different, denoised sonar images may have noise that is not observable to the human eye, thereby reducing the performance of subsequent detection models. At present, some studies have attempted to solve the above problems by fusing denoised images with features extracted from original images that are beneficial to detection, but due to the general quality of denoised images, their performance is still general. Therefore, how to obtain an unsupervised denoising algorithm for underwater sonar images without a synthetic training dataset, so as to obtain higher quality denoised images and enhance detection performance is a technical problem that needs to be solved. Summary of the invention

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a forward-looking sonar image target detection method based on efficient pseudo-labels. The Gaussian mixture model obtained by fitting the original sonar image is used as input to provide the sonar image distribution prior for the deep image prior network, and the best one is selected from the denoised sonar image generated by the deep image prior network as a pseudo-label to help improve the detection performance.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] According to one aspect of the present invention, a method for forward-looking sonar image target detection based on efficient pseudo-labels is provided, comprising the following steps:

[0009] S1. Construct a deep image prior network based on Gaussian mixture model (GMMDIP), take Gaussian mixture model distribution as input, and generate denoised sonar images;

[0010] S2, design a detection-friendly image quality assessment network (DFIQA) to evaluate the quality score of the image and obtain the loss function;

[0011] S3, using the GMMDIP network to generate a denoised image, and using the DFIQA network to evaluate the image quality, and screen out high-quality images as pseudo labels;

[0012] S4, the reconstruction module fuses the backbone network features and performs upsampling to obtain a reconstructed image with the same size as the pseudo label;

[0013] S5. The pseudo-label performs additional constraint optimization on the detection network through MSE loss and detection loss, and the backbone network detects and extracts the features submerged by noise in the original input image.

[0014] As a preferred technical solution, the GMMDIP network includes an up-and-down sampling network reduced to three layers. By fitting the original sonar image with a Gaussian mixture model (GMM), its image distribution characteristic z=GMM(y0) is obtained and used as the input of the GMMDIP network instead of using uniform noise between 0 and 0.1 as input. In this way, the initial distribution prior of each sonar image is pre-provided to the network and combined with the deep prior learned by the GMMDIP network to further improve the denoising performance.

[0015] Use f θ :Z→Y indicates that The parameterized GMMDIP network transforms the tensor / vector z∈Z into the raw sonar image y∈Y. The conditional image distribution p(x * |y0) uses the prior knowledge of the network to restore the denoised image and model it, where the image x * Determined by the measurement y0. Therefore, the final optimization objective expression of GMMDIP is:

[0016]

[0017] This formula can also be expressed as:

[0018]

[0019] The structure of GMMDIP has a jump-connected "encoder-decoder" structure. Since the original DIP network has a five-layer up-and-down sampling network, it takes a long time to achieve the best denoising performance, which increases the time cost of selecting pseudo labels. The present invention reduces the number of layers of the up-and-down sampling network to three layers to minimize the computational burden.

[0020] As a preferred technical solution, the DFIQA network includes two identical shared weight branches, selects denoised images and their entropy image pairs from the detection-friendly image quality assessment network dataset as input, and uses adaptive average pooling to add a fully connected layer after the feature extractor to return the quality score. The quality score obtained by the DFIQA network to evaluate the quality of the reconstructed image will be integrated into the object detection network as a loss function to better optimize the detection performance.

[0021] Furthermore, the input image directly passes through the feature extractor, and the entropy image of the input image first passes through the entropy image attention module (EIA) and then passes through the feature extractor. The EIA is used to obtain the spatial attention weight SA, guiding the input image to pay more attention to the information part related to the detection task. The expression of the attention weight SA is:

[0022]

[0023] in is the entropy image of the input image; H is the height of the image; W is the width of the image; the (1-EI) term is used to adjust the weight of the entropy image, assigning higher weights to the parts of the original image that contain information about the object and lower weights to the remaining areas; σ(·) is the Sigmoid function; δ(·) is the ReLU function; * is the convolution operation; W is the weight parameter of the convolution; b is the bias of the convolution. Use the Sigmoid function to map the values ​​between 0 and 1.

[0024] In the training phase, given two input pairs and And two truth values ​​Q1 and Q2, the expression of the output quality score is:

[0025]

[0026] Among them, θ is the network parameter.

[0027] The loss function expression of DFIQA is:

[0028] L DFIQA =L1Loss(q1,Q1)+L1Loss(q2,Q2)+λL(q1;q2),

[0029] Where λ is a hyperparameter; q is the output quality score; L(q1;q2) is the edge ranking loss; Q is the true value. The expression of edge ranking loss is:

[0030]

[0031] Where ∈ is a hyperparameter.

[0032] As a preferred technical solution, in order to enable the detection backbone network to be optimized by pseudo labels, the backbone features are reconstructed into images through a reconstruction module. Since the first three layers of the backbone network have rich spatial and semantic information, which is helpful for reconstructing images, the first three layers of the backbone network are selected to fuse the features of the three layers through convolutional layers and average pooling layers, and then upsampled to obtain a reconstructed image of the same size as the pseudo label, thereby achieving the effect of additional constrained optimization of the detection network.

[0033] As a preferred technical solution, the expression of the overall detection network loss function is:

[0034] L=L cls +L loc +λL QL +L MSE ,

[0035] Among them, L cls Focal Loss; L loc is Huber Loss, λ is a hyperparameter; L QL Indicates image quality loss; L MSE is the mean square loss;

[0036] The L QL The expression is:

[0037] L QL =sigmoid(-Q(RI i ));

[0038] Among them, RI i To reconstruct the image, Q(RI i ) is the quality score of the reconstructed image.

[0039] Furthermore, the DFIQA network also provides a quality score for the reconstructed image, which is used as the loss function value after passing through the sigmoid function, so that the feature extractor can be fine-tuned on the sonar dataset and the detection network can be jointly optimized.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] (1) Improved detection performance: The introduction of efficient pseudo-labels enables the detection network to more accurately extract features from images, thereby improving the performance of target detection.

[0042] (2) Collaborative optimization of denoising and detection: By combining the GMMDIP network and the DFIQA network, the collaborative optimization of denoising and detection is achieved, which not only removes the noise in the sonar image with high quality, but also retains the features that are beneficial to detection.

[0043] (3) Wide scope of application: The method of the present invention has been trained and tested on multiple real sonar datasets, showing good denoising ability and is suitable for underwater sonar image target detection in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flowchart of the forward-looking sonar image target detection method based on efficient pseudo-labels;

[0045] Figure 2 It is a schematic diagram of the structure of the GMMDIP network;

[0046] Figure 3 It is a schematic diagram of the structure of the DFIQA network;

[0047] Figure 4 Schematic diagram of the structure of the EIA entropy image attention module in the DFIQA network;

[0048] Figure 5 It is a schematic diagram of the structure of the reconstruction module;

[0049] Figure 6 Some examples of images in the MDFD dataset;

[0050] Figure 7 Some examples of images in the UATD dataset;

[0051] Figure 8 This is a visualization of the test results on the MDFD dataset. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0053] The ocean covers most of the earth's surface and contains rich biological, mineral and chemical resources. It is also an important channel for energy transportation and information transmission. Therefore, the development and utilization of marine resources are receiving more and more attention. Underwater robots equipped with sonar equipment are widely used in underwater target detection, marine salvage, underwater navigation and other tasks, which can assist people in the development and utilization of marine resources. Among them, the use of sonar images for target detection has received more and more attention, because it can not only determine the type and location of the target, but also help the operator better understand the underwater scene and reduce the manual burden.

[0054] like Figure 1 As shown, the present invention discloses a forward-looking sonar image target detection method based on efficient pseudo-labels, comprising the following steps:

[0055] S1, build the GMMDIP network, take the Gaussian mixture model distribution as input, and generate denoised sonar images;

[0056] S2. Design the DFIQA network to evaluate the quality score of the image and obtain the loss function;

[0057] S3, using the GMMDIP network to generate denoised images, and using the DFIQA network to evaluate image quality, and select high-quality images as pseudo labels;

[0058] S4, the reconstruction module fuses the backbone network features and performs upsampling to obtain a reconstructed image with the same size as the pseudo label;

[0059] S5. Pseudo-labeling uses MSE loss and detection loss to perform additional constraint optimization on the detection network. The backbone network detects and extracts features that are submerged in noise in the original input image.

[0060] The GMMDIP network includes an up- and down-sampling network reduced to three layers. The image distribution characteristic z=GMM(y0) is obtained by performing GMM fitting on the original sonar image and used as the input of the GMMDIP network instead of using uniform noise between 0 and 0.1 as input. This allows the initial distribution prior of each sonar image to be provided to the network in advance and combined with the deep prior learned by the GMMDIP network to further improve the denoising performance. Figure 2 The figure shows the schematic diagram of the structure of the GMMDIP network.

[0061] Use f θ :Z→Y indicates that The parameterized GMMDIP network transforms the tensor / vector z∈Z into the raw sonar image y∈Y. The conditional image distribution p(x * |y0) uses the prior knowledge of the network to restore the denoised image and model it, where the image x *Determined by the measurement y0. Therefore, the final optimization objective expression of GMMDIP is:

[0062]

[0063] This formula can also be expressed as:

[0064]

[0065] The structure of GMMDIP has an "encoder-decoder" structure with jump connections. Since the original DIP network has a five-layer up-and-down sampling network, it takes a long time to achieve the best denoising performance, which increases the time cost of selecting pseudo labels. The present invention reduces the number of layers of the up-and-down sampling network to three layers to minimize the computational burden. In addition, LeakyReLU is used as the nonlinear activation function, the stride is set to 2 in the convolution layer for downsampling, and the nearest neighbor interpolation technique is used for upsampling. In order to minimize the artifacts introduced by padding, reflection padding is used as the zero padding technique. To further speed up the convergence of the network, the convolution parameters of the first three shallow convolutional layers and the last three convolutional layers are frozen after the network is randomly initialized. In addition, we set the first layer of GMMDIP to batch normalization to make the optimization function smoother and allow batch normalization to be optimized during training.

[0066] like Figure 3 As shown in the figure, it is a schematic diagram of the structure of the DFIQA network. The DFIQA network consists of two identical shared weight branches. It selects denoised images and their entropy image pairs from the detection-friendly image quality assessment network dataset as input, and uses adaptive average pooling to add a fully connected layer after the feature extractor to return the quality score. The quality score obtained by the DFIQA network to evaluate the quality of the reconstructed image will be integrated into the target detection network as a loss function to better optimize the detection performance. The input image passes directly through the feature extractor, and the entropy image of the input image passes through EIA first and then through the feature extractor. EIA is used to obtain the spatial attention weight SA to guide the input image to pay more attention to the information part related to the detection task. The EIA structure is shown in Figure 4 The expression of the attention weight SA is shown as:

[0067]

[0068] in is the entropy image of the input image; H is the height of the image; W is the width of the image; the (1-EI) term is used to adjust the weight of the entropy image, assigning higher weights to the parts of the original image that contain information about the object and lower weights to the remaining areas; σ(·) is the Sigmoid function; δ(·) is the ReLU function; * is the convolution operation; W is the weight parameter of the convolution; b is the bias of the convolution. Use the Sigmoid function to map the values ​​between 0 and 1.

[0069] In the training phase, given two input pairs and And two truth values ​​Q1 and Q2, the expression of the output quality score is:

[0070]

[0071] Among them, θ is the network parameter.

[0072] The loss function expression of DFIQA is:

[0073] L DFIQA =L1Loss(q1,Q1)+L1Loss(q2,Q2)+λL(q1;q2),

[0074] Where λ is a hyperparameter; q is the output quality score; L(q1;q2) is the edge ranking loss; Q is the true value. The expression of edge ranking loss is:

[0075]

[0076] Where ∈ is a hyperparameter and is set to 0.5. When actually testing the DFIQA network, only one branch of the DFIQA network is selected to predict image quality.

[0077] In order to make the detection backbone network optimized by pseudo labels, the backbone features are reconstructed into images through the reconstruction module, such as Figure 5 As shown in the figure. Since the first three layers of the backbone network have rich spatial and semantic information, which is helpful for reconstructing images, the first three layers of the backbone network are selected to fuse the features of the three layers through convolutional layers and average pooling layers, and then upsample to obtain a reconstructed image with the same size as the pseudo-label, so as to achieve the effect of additional constraint optimization of the detection network. The reconstruction module serves as a medium, allowing the pseudo-label to optimize the entire detection network through MSE loss and detection loss, so that the backbone network can extract features that are submerged by noise in the original input image.

[0078] In order to effectively optimize the detection network, the weights of the feature extraction network are frozen in the first 50 epochs, and the Resnet50 network is used for pre-training on the COCO dataset. Then the network is unfrozen in the last 50 epochs to optimize the feature extractor and the back-end detection head network, so that the feature extractor can be fine-tuned on the sonar dataset, thereby improving the overall detection performance. The overall detection network loss function expression of the present invention is:

[0079] L=L cls +L loc +λLQL +L MSE ,

[0080] Among them, L cls Focal Loss; L loc is Huber Loss, λ is a hyperparameter; L QL Indicates image quality loss; L MSE is the mean square loss;

[0081] L QL The expression is:

[0082] L QL =sigmoid(-Q(RI i ));

[0083] Among them, RI i To reconstruct the image, Q(RI i ) is the quality score of the reconstructed image.

[0084] L MSE L2Loss, used to ensure RI i and the pseudo-label PL selected by DFIQA i The consistency between them is expressed as:

[0085]

[0086] Where |Ω| is the area of ​​the graph Ω.

[0087] Furthermore, the DFIQA network also provides a quality score for the reconstructed image, which is used as the loss function value after passing through the sigmoid function, so that the feature extractor can be fine-tuned on the sonar dataset and the detection network can be jointly optimized.

[0088] The present invention is trained and tested on two real sonar datasets to verify its effectiveness, namely the Marine Debris Forward-Looking Sonar Dataset (MDFD) and the Underwater Acoustic Target Detection (UATD) dataset. Examples of images of the two datasets are shown in Figure 2. Figure 6 and Figure 7As shown. The MDFD dataset was captured by the Marine Systems Laboratory (Heriot-Watt University) using the ARIS Explorer3000 forward-looking sonar with a frequency of 3.0MHz, which contains 1868 sonar images of 11 categories, including bottles, cans, chains, beverage cans, hooks, propellers, shampoo bottles, upright bottles, tires, and valves. The UATD dataset was captured by the Tritech 1200ik forward-looking sonar located in Weihai, China, and contains 9200 sonar images of 10 categories, including cubes, balls, cylinders, human bodies, tires, round cages, square cages, metal barrels, airplanes, and remote-controlled robots. For the MDFD dataset, the present invention allocates 1681 images for training and 187 images for testing the detection model performance. For the UATD dataset, the present invention allocates 8280 images for training and 920 for testing the detection model performance.

[0089] The experimental environment was built with the Pytorch 1.2.0 training framework, and the training was performed on two NVIDIA GeForce RTX 2080SUPER GPUs with 8GB of memory. The GMMDIP network was optimized using the Adam optimizer, with 900 iterations per image and an initial learning rate of 1e-1. The DFIQA network was trained for 200 epochs using the Adam optimizer, with an initial learning rate of 2e-5, a weight decay of 5e-4, and a batch size of 16. The overall invention method was first trained 50 times using the Adam optimizer, with an initial learning rate of 1e-4 and a batch size of 4. Subsequently, another 50 trainings were performed with an initial learning rate of 1e-5 and a batch size of 2. The overall training process used a linear learning rate decay of 0.96.

[0090] Currently, image quality assessment algorithms mainly prioritize human perception and ignore quality assessment methods for machine vision, especially in detection tasks. The differences in the characteristics of machine vision and human vision may make evaluation criteria specific to human vision unsuitable for machine vision tasks. In addition, there are few dedicated datasets for machine vision, and building the DFIQA dataset is crucial for training a quality assessment network DFIQA that can generate detection-friendly labels.

[0091] The method for constructing DFIQA is as follows: GMMDIP first generates a set of pseudo labels for 900 iterations, that is, starting from the 100th iteration, a denoised sonar image is selected as a pseudo label every 100 intervals, thereby generating 9 denoised sonar images for each sonar image. In addition, a corresponding original image is added to each sonar image as a pseudo label, thereby generating 10 sets of pseudo labels as the DFIQA dataset. In addition, in order to enhance the generalization ability of the dataset, the 10 sets of pseudo labels are applied to three single-stage detection networks, YOLOv3, SSD and RetinaNet. To evaluate these three detection networks, two evaluation indicators commonly used in target detection tasks, the product of IoU and AP, are introduced, and the definition formula is:

[0092] MOS DFIQA =AP×IoU(Ω g ,Ω p ),

[0093] Among them, Ω g is the sample frame area, Ω p is the predicted box area. In addition, the test results of the three different detection frameworks are weighted and averaged to obtain a more generalized quality evaluation value. Finally, the DFIQA training dataset corresponding to the MDFD dataset was constructed with 18,680 images, and the DFIQA training dataset corresponding to the UATD dataset was constructed with 92,000 images, representing the performance of the detection models driven by different pseudo labels.

[0094] The Spearman rank correlation coefficient (SROCC) and Pearson linear correlation coefficient (PLCC) were used to evaluate the performance of DFIQA. SROCC evaluates the monotonicity of the method, while PLCC evaluates the accuracy of the method. Generally, better methods are expected to produce higher indicator values.

[0095] Invention method detection performance evaluation index: Use the average precision (AP) in COCO, which is a common performance indicator in object detection. AP is the average of the average precision of all categories under 10 IoU thresholds, and the expression is:

[0096]

[0097] In addition, AP 50 and AP 75 It is also used as a metric. AP is calculated when the predicted bounding box overlaps with the ground truth by at least 50%. 50 AP is measured, and when the overlap is at least 75%, AP 75 Measure APs.

[0098] The effects of denoised images generated by two GMMDIP models on detection performance are compared, one of which takes GMM as input and the other takes uniform noise between 0 and 0.1 as input. Experimental results show that the overall AP performance of GMMDIP using GMM as input on the MDFD and UATD datasets reaches 67.2% and 38.4% respectively, which is better than the detection performance of 65.5% and 36.9% when using uniform noise as input.

[0099] The performance of the DFIQA network will affect the quality of the selected pseudo-labels, thereby affecting the overall detection performance. The effectiveness of the DFIQA network is verified using the two indicators SROCC and PLCC. By training the DFIQA network using the two created DFIQA datasets, the SROCC indicator reached 78.33% and the PLCC indicator reached 64.39% on the MDFD dataset; the SROCC indicator reached 62.42% and the PLCC indicator reached 61.62% on the UATD dataset. The objective effectiveness of the DFIQA network can be verified, and then the detection performance will be verified.

[0100] It is compared with 9 advanced detection methods on the MDFD and UATD datasets. Figure 8 It can be concluded that the invented method is significantly better than the other nine sonar image target detection methods in terms of AP results on the MDFD dataset, and achieves an AP value of 67.8%, which is 0.9% higher than the second value.

[0101] Table 1 Performance comparison of the present invention on the MDFD dataset

[0102]

[0103]

[0104] For the UATD dataset, from Figure 7 It can be seen that the visual effect of the UATD dataset is not good and it is difficult to distinguish the information in the image. Therefore, only quantitative experimental results are given, as shown in Table 2. The inventive method achieves an AP value of 39.8% on the UATD dataset, which is 1.8% higher than the second value. Compared with the benchmark method model RetinaNet, the present invention improves the AP values ​​of 2.9% and 3.5% on the MDFD dataset and UATD dataset, respectively. At the same time, the present invention has a significant improvement in AP 50 and AP 75 The results also showed excellent performance, and the accuracy of target positioning was improved.

[0105] Table 2 Performance comparison of the invented method on the UATD dataset

[0106]

[0107] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A forward-looking sonar image target detection method based on efficient pseudo-labeling, characterized in that: The following steps are involved: S1. Construct a deep image prior network based on Gaussian mixture model, take Gaussian mixture model distribution as input, and generate denoised sonar images; S2. Design a detection-friendly image quality assessment network to evaluate the quality score of the image and obtain the loss function; S3, using the prior network to generate a denoised image, and using the detection-friendly image quality assessment network to evaluate the image quality, and screening out high-quality images as pseudo labels; S4, the reconstruction module fuses the backbone network features and performs upsampling to obtain a reconstructed image with the same size as the pseudo label; S5. The pseudo-label performs additional constraint optimization on the detection network through MSE loss and detection loss, and the backbone network detects and extracts the features submerged by noise in the original input image.

2. The method for forward-looking sonar image target detection based on efficient pseudo-labeling according to claim 1, characterized in that: The deep image prior network based on the Gaussian mixture model includes a three-layer up- and down-sampling network, and uses the Gaussian mixture model obtained by fitting the original sonar image as input to provide a priori of the sonar image distribution.

3. The method for forward-looking sonar image target detection based on efficient pseudo-labeling according to claim 2 is characterized in that: The final optimization target expression of the deep image prior network based on Gaussian mixture model is: Among them, f θ :Z→Y Parameterized deep image prior network based on Gaussian mixture model, z∈Z is a tensor / vector, y∈Y is the original sonar image.

4. The method for forward-looking sonar image target detection based on efficient pseudo-labeling according to claim 1, characterized in that: The detection-friendly image quality assessment network includes two identical shared-weight branches, selects denoised images and their entropy image pairs from the detection-friendly image quality assessment network dataset as input, and uses adaptive average pooling to add a fully connected layer after the feature extractor to return a quality score.

5. The method for forward-looking sonar image target detection based on efficient pseudo-labeling according to claim 4 is characterized in that: The input image directly passes through the feature extractor, and the entropy image of the input image first passes through the entropy image attention module and then passes through the feature extractor.

6. The method for forward-looking sonar image target detection based on efficient pseudo-labeling according to claim 5 is characterized in that: The entropy image attention module is used to obtain the spatial attention weight SA, which guides the input image to pay more attention to the information part related to the detection task. The expression of the attention weight SA is: in is the entropy image of the input image; H is the height of the image; W is the width of the image; the (1-EI) term is used to adjust the weight of the entropy image, assigning higher weights to the parts of the original image that contain information about the object and lower weights to the remaining areas; σ(·) is the Sigmoid function; δ(·) is the ReLU function; * is the convolution operation; W is the weight parameter of the convolution; b is the bias of the convolution.

7. The method for forward-looking sonar image target detection based on efficient pseudo-labeling according to claim 1, characterized in that: The loss function expression of the designed detection-friendly image quality assessment network is: L DFIQA =L1Loss(q1,Q1)+L1Loss(q2,Q2)+λL(q1;q2), Where λ is a hyperparameter; q is the output quality score; L(q1;q2) is the edge ranking loss; Q is the true value.

8. The method for forward-looking sonar image target detection based on efficient pseudo-labeling according to claim 1, characterized in that: The reconstruction module selects the first three layers of the backbone network and fuses the features of the three layers through convolutional layers and average pooling layers.

9. The method for forward-looking sonar image target detection based on efficient pseudo-labels according to claim 1, characterized in that: The expression of the overall detection network loss function is: L=L cls +L loc +λL QL +L MSE , Among them, L cls Focal Loss; L loc is Huber Loss, λ is a hyperparameter; L QL Indicates image quality loss; L MsE is the mean square loss; The L QL The expression is: L QL =sigmoid(-Q(RI i )), Among them, RI i To reconstruct the image, Q(RI i ) is the quality score of the reconstructed image.

10. The method for forward-looking sonar image target detection based on efficient pseudo-labeling according to claim 1, characterized in that: The detection-friendly image quality assessment network also provides a quality score for the reconstructed image, which is used as a loss function value after passing through a sigmoid function to jointly optimize the detection network.

Images