Underwater complex target detection method and system based on generation and detection collaborative optimization

Abstract

The invention belongs to the technical field of underwater target detection, and discloses an underwater complex target detection method and system based on generation and detection collaborative optimization, and the method comprises the steps: generating a side-scan sonar simulation sample through employing a CycleGAN, and screening out a high-quality sample through employing a pre-trained YOLO model, so as to construct an enhanced data set; an improved target detection model with YOLOv8 as a baseline is constructed, the model integrates a lightweight backbone network formed by deep separable convolution and an inverted bottleneck structure, a feature fusion neck network adopting element multiplication and residual connection, and an EMA attention module introduced in front of a detection head, and the model is trained by using the enhanced training data set; and performing target detection on an input underwater side-scan sonar image by using the improved YOLOv8 model after training is completed. According to the method, the problems of scarcity of underwater samples and multi-scale detection of targets are effectively solved, and the model lightweight is realized while the detection precision is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of underwater target detection, and particularly relates to an underwater complex target detection method and system based on generation and detection collaborative optimization. BACKGROUND

[0002] In recent years, with the rapid growth of the demand for marine environment observation, marine engineering development and marine emergency search and rescue, high-resolution seabed acoustic detection technologies represented by multi-beam and side scan sonar have developed rapidly, and efficient and accurate underwater target detection has become a key technology in marine detection and decision support. At present, most underwater target detection methods are based on acoustic, magnetic, optical or electrical detection. Sound waves have become the most widely used underwater target detection method due to their ability to propagate over long distances in water. Among acoustic methods, side scan sonar (SSS) is widely used for detecting underwater targets. As a kind of high-frequency towed sonar, it has the advantages of low cost and high imaging resolution, and has become an ideal solution for large-scale marine detection.

[0003] Early side scan sonar image target detection mainly relies on manual interpretation, and the efficiency and accuracy are limited by personnel experience. Traditional machine learning methods achieve target recognition through image preprocessing, artificial feature extraction (such as texture, edge, etc.) and classification algorithms (such as SVM, PCA, K-means, etc.). Although certain effects have been achieved in specific scenarios, such methods have complex processes, low efficiency, and the preprocessing, feature extraction and classification stages are independent of each other, lack of end-to-end optimization, are prone to error accumulation, and are difficult to meet the needs of high real-time, high precision and strong robustness of marine detection.

[0004] To obtain high-quality labeled samples, data augmentation methods are widely used. Traditional data augmentation methods mainly include geometric transformation (such as flipping, rotation and scaling), color adjustment (such as brightness and contrast disturbance) and pixel value disturbance (such as adding Gaussian noise) and so on. However, traditional data augmentation can only perform linear transformation in the image space of the original data set. These methods essentially make small disturbances and copies within the neighborhood of existing samples. Although they can increase the amount of data to some extent, they do not introduce any new information or new semantics beyond the distribution range of the original data set, and it is difficult to fundamentally improve the model generalization ability. In recent years, as a deep learning-based generative model, the generative adversarial network (GAN) proposed by Ian J. Goodfellow et al. has been used to generate simulated images to expand data samples in multiple fields. The deep convolutional generative adversarial network (DCGAN) proposed by Alec Radfor creatively combines convolutional neural networks (CNN) and GAN, and significantly improves the training stability by imposing a series of constraints on the network structure, becoming the basis for subsequent improved GAN models. The cycle-consistent generative adversarial network (CycleGAN) proposed by Zhu et al. further breaks through the limitation of relying on paired data for cross-domain image conversion, and realizes the bidirectional mapping between the source domain and the target domain by constructing a double-adversarial learning framework containing two sets of generators and discriminators. Liu et al. developed an acoustic image simulator based on 3D modeling software, and used the simulated data set to generate realistic acoustic images with the help of CycleGAN. This method is simple to operate and can generate effective and realistic acoustic data. Zheng et al. proposed a small sample target detection method based on side-scan sonar, which combines CycleGAN and YOLOv8, effectively alleviates the problem of sample scarcity of side-scan sonar, and improves the detection accuracy, verifying the feasibility of CycleGAN in generating side-scan sonar simulation samples.

[0005] Due to the diversity of actual seabed environment target categories and other problems, it is difficult for side-scan sonar target detection models to extract sufficient effective features during learning. In recent years, many domestic and foreign researchers have conducted extensive research on this problem. Yu et al. proposed a target detection model that integrates transformer modules and attention mechanisms into YOLOv5s to address the sparsity of side-scan sonar image targets, and improved the detection accuracy by 12.5%. Cao et al. proposed a lightweight algorithm for side-scan sonar image target detection based on multi-scale feature fusion and attention mechanisms, which enhanced the model's ability to fuse and represent multi-scale features while improving the detection accuracy by 9.9%. Wen et al. introduced a feature scaling factor into the YOLOv7 model to address the uncertainty of geometric features in side-scan sonar target features, which improved the detection accuracy by 9.28%. YOLOv10 and YOLOv11 improved upon YOLOv8, resulting in a lightweight and high-performance model that achieved good detection accuracy in side-scan sonar target detection tasks.

[0006] Although numerous studies have demonstrated the effectiveness of GAN in sonar image generation, existing methods still face three major challenges: first, the quality of generated simulation images lacks stability, second, there is a lack of systematic evaluation of their actual contribution to downstream target detection tasks, making it difficult to effectively improve detection network performance, and third, existing research methods typically treat simulation sample generation and target detection tasks as two independent modules for processing. Such separate optimization strategies have obvious limitations: on the one hand, the generated simulation samples often fail to fully meet the feature requirements of the target detection model, resulting in suboptimal sample quality and difficulty in effectively improving detection performance; on the other hand, the simulation sample generation model itself lacks feedback signals from the detection task, which restricts its optimization process, forming a two-way performance bottleneck. Therefore, breaking through the traditional paradigm of independent generation and detection models and establishing a new strategy for their collaborative optimization is crucial for improving detection performance. Second, although existing advanced detection methods perform well in side-scan sonar image target detection, they struggle to balance lightweight design, small sample learning ability, and multi-scale feature extraction and fusion for underwater heterogeneous targets. In this context, developing a target detection method that is both practical and adaptable is essential for promoting the deployment and application of side-scan sonar target detection technology in actual engineering.

[0007] Through the above analysis, the problems and defects of the prior art are: (1) Early side-scan sonar image target detection mainly relies on manual interpretation, and the efficiency and accuracy are limited by personnel experience; traditional machine learning methods have complex processes, low efficiency, and lack of end-to-end optimization, which can easily lead to error accumulation and make it difficult to meet the high real-time, high-precision, and strong robustness requirements of marine exploration.

[0008] (2) The simulation samples generated by the existing separate optimization strategy fail to fully meet the feature requirements of the target detection model, resulting in unsatisfactory sample quality and difficulty in effectively improving detection performance; the simulation sample generation model itself lacks feedback signals from the detection task, which restricts the optimization process and forms a two-way performance bottleneck.

[0009] (3) Existing underwater target detection methods are difficult to simultaneously achieve lightweight design, small sample learning ability, and multi-scale feature extraction and fusion for heterogeneous underwater targets. Summary of the Invention

[0010] To overcome the problems existing in related technologies, the present invention discloses an underwater complex target detection method and system based on the collaborative optimization of generation and detection, particularly relating to an underwater complex target detection method and system that takes into account both high-quality simulation sample enhancement and target scale differences. The technical solution is as follows: This invention is implemented as follows: A method for detecting complex underwater targets based on collaborative optimization of generation and detection includes the following steps: S1. Simulation Sample Generation and Screening: CycleGAN is used to generate side-scan sonar simulation samples in the target domain from the source domain image. The generated simulation samples are screened using a pre-trained target detection model, and samples with confidence scores higher than a preset threshold are retained as high-quality simulation samples. S2. Training Dataset Construction: The high-quality simulation samples are fused with the original side-scan sonar dataset to construct an enhanced training dataset; S3. Detection Model Improvement and Training: Construct an improved object detection model based on YOLOv8 as the baseline, and train the improved YOLOv8 model using the enhanced training dataset. The improved YOLOv8 model includes: A lightweight feature extraction backbone network based on depthwise separable convolution and a general inverted bottleneck structure; Lightweight feature fusion of the neck network using element-wise multiplication and residual connections; and An efficient multi-scale attention mechanism (EMA) module is introduced in front of the detection head; S4. Target Detection Inference: Target detection is performed on the input underwater side-scan sonar image using the improved YOLOv8 model that has been trained.

[0011] In step S1, the source domain image is a remote sensing optical image, and the target domain is a side-scan sonar image; the pre-trained target detection model is a YOLOv8 model, and the preset confidence threshold is 0.5.

[0012] In step S1, after the simulation sample screening, a manual verification step is also included to remove false and missed samples and to perform data augmentation on the valid samples.

[0013] In step S3, the lightweight feature extraction backbone network includes: Depthwise separable convolution consists of depthwise convolution and pointwise convolution. Depthwise convolution uses a single-channel convolution kernel to independently process the spatial dimension of each input channel, extracting local features while keeping the number of channels constant. Pointwise convolution uses a 1×1 convolution kernel to fuse cross-channel information and freely control the number of output feature maps. The general inverted bottleneck structure includes four configurable stages: an initial depthwise convolution to process spatial features, a 1×1 expanding convolution to dynamically increase the channel dimension through the expansion ratio, an intermediate depthwise convolution to enhance feature extraction, and a 1×1 projective convolution to compress channels and achieve dimensionality reduction; through parameterized design, the kernel size of each stage can be customized, supporting 3×3, 5×5 or skip operations, and supporting flexible downsampling control strategies. All convolutional layers employ the ReLU6 activation function and batch normalization.

[0014] Furthermore, the feature extraction process of the lightweight feature extraction backbone network is as follows: the input image first goes through a primary feature construction stage to generate multiple sets of basic feature representations; then, these features are integrated and enhanced through a feature interaction and fusion module to produce higher-order features that are richer in information and more discriminative; then, a deep refinement stage purifies these features; finally, through a multi-scale context information aggregation layer, context information of different ranges is adaptively captured and fused to output a final feature representation with highly enriched information, laying the foundation for subsequent decoding and prediction tasks; The activation function is an improved version of the ReLU activation function. For negative input values, it restricts the output to zero or higher, and for positive input values, it restricts it to 6 or lower, thus producing an output range of [0, 6]. The expression is: ; In the formula, This is an operation to retrieve the minimum value within the parentheses. This is an operation to retrieve the maximum value within the parentheses. The representative model is the original linear value output by the width and height offset of the bounding box.

[0015] In step S3, the lightweight feature fusion neck network employs a combination of element-wise multiplication and residual connections. Through a dual-path feature extraction mechanism, one feature path is activated by ReLU6 and then element-wise multiplied with the other path, achieving dynamic feature selection. This adaptively enhances target-related features and suppresses background noise. Introducing convolutions at a specific depth expands the model's receptive field, enhancing its ability to perceive target structures and complex textures. Element-wise multiplication refers to fusing two input feature vectors through element-wise multiplication (i.e., element-by-element multiplication), thereby achieving high-dimensional nonlinear feature mapping in a low-dimensional input space. We define... (The sentence is incomplete and requires further context to translate accurately.) , and The element-wise multiplication formula is shown below: ; In the formula, and These are combinations of two different weight matrices and biases, specifically represented as follows: , This is the weight matrix. For bias terms, This is a matrix transpose operation. For the set of real numbers, Input feature vector The corresponding number in The components of each channel, For input vectors The corresponding number in The components of each channel, For the input feature vector, The channel dimension of the input feature. All are channel indexes.

[0016] In step S3, the efficient multi-scale attention mechanism (EMA) module processes the characteristics of underwater side-scan sonar images. This mechanism divides the input feature map into G sub-features in the channel dimension through channel grouping and dimension reshaping strategies to adapt to the complex distribution of target multi-scale coexistence and background clutter in sonar images. Based on this, a dual-branch parallel processing structure is adopted: the 1×1 convolution branch incorporates a coordinate attention mechanism, which aims to effectively capture the spatial contextual associations along the track and perpendicular to the track in underwater side-scan sonar images, enhancing the model's ability to perceive blurred, low-contrast acoustic targets; the 3×3 convolution branch focuses on enhancing the extraction of local details and texture features in sonar images to suppress acoustic noise interference, jointly improving the detection accuracy and robustness in complex underwater acoustic environments.

[0017] In step S4, the trained improved YOLOv8 model is used to detect multi-scale targets in underwater side-scan sonar images, including drowning people, aircraft, and ships.

[0018] Another objective of this invention is to provide an underwater complex target detection system based on co-optimization of generation and detection. This system is used to implement the aforementioned underwater complex target detection method, comprising: The simulation sample generation and screening unit uses CycleGAN to generate side-scan sonar simulation samples in the target domain from source domain images. It uses a pre-trained target detection model to screen the generated simulation samples, retaining samples with confidence scores higher than a preset threshold as high-quality simulation samples. The training dataset construction unit fuses the high-quality simulation samples with the original side-scan sonar dataset to construct an enhanced training dataset. The detection model improvement and training unit constructs an improved object detection model based on YOLOv8 as a baseline, and trains the improved YOLOv8 model using the enhanced training dataset. The target detection inference unit uses the trained improved YOLOv8 model to perform target detection on the input underwater side-scan sonar image.

[0019] Another objective of this invention is to provide an application of an underwater complex target detection method in marine environmental observation, marine engineering development, and maritime emergency search and rescue scenarios.

[0020] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention constructs a task-oriented simulation sample generation and screening mechanism and improves the multi-scale feature fusion capability of the target detection network. While ensuring the model's lightweight nature, it effectively enhances the model's detection performance for underwater heterogeneous targets in small-sample scenarios, providing a feasible technical solution for practical engineering applications. This invention addresses the problem of a lack of high-quality labeled samples in side-scan sonar target detection tasks and alleviates the poor feature extraction and fusion capabilities of YOLOv8 for underwater heterogeneous targets.

[0021] Secondly, this invention empowers the marine economy and pioneers a new paradigm for precise underwater detection: This system can run in real time on edge computing devices (such as the onboard computer of an autonomous underwater vehicle, AUV), achieving a unity of "precise perception" and "instant response," providing core technical support for the autonomous decision-making of unmanned platforms. After migration, it can provide cost-effective and highly reliable automated detection tools for fields such as underwater oil and gas resource exploration, subsea pipeline inspection, underwater archaeology, and fisheries resource surveys. By improving the level of automation in detection, the cost of manual verification can be significantly reduced, and operational efficiency and safety can be improved, which is expected to save related enterprises more than 30% of their operating costs.

[0022] Third, this invention breaks through the traditional paradigm of separating the "sample generation model" and the "target detection model" in existing underwater sonar target detection technologies. Existing solutions either rely on limited real data, resulting in insufficient model generalization ability, or simply introduce simulation data, leading to model performance degradation. Addressing this core bottleneck, this invention, for the first time, constructs a task-oriented simulation sample generation and screening mechanism. This system, through the interactive feedback between generative adversarial networks and the target detection model, achieves the effective utilization of high-quality simulation samples, fundamentally solving the industry problem of data scarcity and model overfitting, and providing a completely new technical path for high-precision underwater target detection under small sample conditions.

[0023] Fourth, this invention overcomes the technical bias that "simulation data generated by Generative Adversarial Networks (GANs) will harm model performance." Instead of simply "using" simulation data, it creatively introduces a screening mechanism that uses a pre-trained object detection model as the "judge," selecting only simulation samples that the current model can identify with high confidence. Through collaborative optimization, it directionally strengthens the object detection model's ability to identify difficult samples. This overcomes the technical bias that "lightweight models inevitably sacrifice accuracy." This invention effectively improves the average accuracy of the object detection model while significantly reducing the number of model parameters and weights. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a flowchart of the underwater complex target detection method based on collaborative optimization of generation and detection provided in the embodiments of the present invention; Figure 2 This is a flowchart of the simulation sample generation and screening technology provided in the embodiments of the present invention; Figure 3 This is an improved YOLOv8 network structure diagram provided in an embodiment of the present invention; Figure 4 This is a flowchart of the depth-separable convolution processing provided in an embodiment of the present invention; wherein, depthwise convolution is performed first, followed by pointwise convolution; Figure 5 This is a diagram of a general inverted bottleneck network structure provided in an embodiment of the present invention; Figure 6 This is a structural diagram of the lightweight feature extraction module provided in an embodiment of the present invention; Figure 7 This is a structural diagram of the lightweight feature fusion module provided in an embodiment of the present invention; Figure 8This is a structural diagram of the EMA attention mechanism provided in an embodiment of the present invention; Figure 9 This is a technical flowchart of the underwater complex target detection method that takes into account the enhancement of high-quality simulation samples and the differences in target scale provided by the embodiments of the present invention; Figure 10 This is a graph showing the average accuracy (mAP@50) of YOLOv8s and the method of the present invention during the training process, provided in the embodiments of the present invention. Figure 11 This is a comparison diagram of the detection results of YOLOv8s provided in the embodiments of the present invention and the side-scan sonar images of the present invention method; Figure 12 This is a comparison image of the receptive field and heat map of YOLOv8s and the method of the present invention under the same image provided in the embodiments of the present invention.

[0025] Figure 13 This is an example diagram illustrating the implementation of conventional data augmentation in this invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0027] The innovation of this invention lies in: 1. Considering the scarcity of side-scan sonar target samples and the poor image quality generated by CycleGAN, a side-scan sonar image target detection method with co-optimization of generation and detection is proposed by effectively combining CycleGAN and YOLO. This method ensures that the synthesized samples have both visual realism and meet the feature requirements of downstream detection tasks, effectively improving the usability and task relevance of the generated data.

[0028] 2. To improve YOLOv8's ability to extract local and global features while reducing the model's complexity and size, a lightweight feature extraction module is proposed.

[0029] 3. To effectively enhance YOLOv8's ability to fuse multi-scale spatial features, a lightweight feature fusion module is proposed.

[0030] Example 1, such as Figure 1 As shown, the underwater complex target detection method based on collaborative optimization of generation and detection provided in this embodiment of the invention includes the following steps: S1. Simulation Sample Generation and Screening: CycleGAN is used to generate side-scan sonar simulation samples in the target domain from the source domain image. The generated simulation samples are screened using a pre-trained target detection model, and samples with confidence scores higher than a preset threshold are retained as high-quality simulation samples. Data-driven deep learning models typically rely on large-scale, high-quality training data to achieve superior performance. However, in practical applications, they often face problems such as data scarcity, inconsistent quality, and class imbalance due to cost and time constraints, which is particularly prominent in side-scan sonar image tasks. Due to the complex underwater environment and acquisition conditions, acquiring and annotating a large number of high-quality images is costly and difficult to implement. To alleviate sample scarcity, data augmentation techniques have been widely introduced. Traditional data augmentation methods (such as geometric, color, and pixel transformations) are simple to operate, but the diversity of generated images is limited, making it difficult to fundamentally improve the model's generalization ability. Furthermore, traditional data augmentation can only perform linear transformations within the image space of the original dataset. These methods essentially perform small perturbations and replications within the neighborhood of existing samples. While they can increase the amount of data to some extent, they do not introduce any new information or semantics beyond the distribution range of the original dataset, making it difficult to fundamentally improve the model's generalization ability. In recent years, Generative Adversarial Networks (GANs) have been widely introduced into the field of sample augmentation. Among them, CycleGAN can learn cross-domain mappings without paired data, generating realistic side-scan sonar images, effectively enhancing data diversity, and thus improving the training effect and generalization ability of target detection models.

[0031] Figure 2This paper demonstrates the overall process of simulation sample generation and selection. The method employs the CycleGAN model, whose core is a symmetric generative adversarial network structure containing two generators (Gx and Gy) and two discriminators (Dx and Dy), enabling unpaired image style transfer from the source domain (remote sensing optical images) to the target domain (side-scan sonar images). To select high-quality simulation samples that positively contribute to the training of the detection model, this invention introduces a pre-trained YOLOv8 model as an evaluation module. The detection confidence obtained from forward inference is used as the criterion for judging sample quality, retaining only samples with confidence scores higher than a preset threshold, thereby ensuring that the selected samples possess good semantic consistency and feature effectiveness. During the screening process, the confidence threshold was set to 0.5, primarily based on the following two considerations: First, mAP@0.5 is a widely adopted performance metric in the field of object detection (such as the COCO evaluation standard). Using this as the screening threshold ensures that the model validation benchmark is consistent with the screening criteria and the model evaluation system, guaranteeing the retention of effective detection results with IoU ≥ 0.5. Second, this threshold helps to achieve a balance between recall and precision, thereby optimizing data utilization and sample quality. Side-scan sonar images typically have sparse targets and complex backgrounds, and the simulated images generated by CycleGAN are prone to artifacts or structural distortions. If the threshold is set too high, although it can improve sample quality, it may lead to too many weak target samples being falsely filtered, reducing the amount of usable data; if the threshold is too low, it is easy to introduce noisy samples, interfering with model training and leading to more false detections. Therefore, a confidence threshold of 0.5 is the engineering optimal solution that balances data utilization (Recall) and quality (Precision), especially suitable for scenarios where generated data is scarce.

[0032] S2. Training Dataset Construction: The high-quality simulation samples are fused with the original side-scan sonar dataset to construct an enhanced training dataset; In the field of target detection, YOLOv8 demonstrates significant advantages in balancing accuracy and speed, architectural simplicity, and a mature community ecosystem, making it a widely adopted benchmark model in industry and research applications. Its anchor-free design simplifies the training process and provides a unified framework to support multiple tasks such as detection and segmentation, laying a solid foundation for customized improvements in specific scenarios. However, the backbone network of the YOLOv8 base model has limited ability to extract small targets and weak features, and its default architecture lacks generalization ability for multi-scale targets, leading to a significant performance degradation in complex underwater scenarios. Although subsequent releases like YOLOv10 further optimized efficiency, and YOLOv11 enhanced versatility, these advanced algorithms either simplified their structures to pursue minimal latency or significantly increased model complexity by introducing multimodal fusion. Considering that underwater side-scan sonar tasks typically have limited data scale and computational resources, and prioritize robustness to specific targets rather than general applicability, this invention uses YOLOv8 as a baseline model for improvement. Figure 3 The improved network structure is shown.

[0033] To address the challenges of large scale variations and feature scarcity in underwater heterogeneous targets, this invention proposes a lightweight feature extraction module to replace the original backbone network. This module combines depthwise separable convolution with a general inverted bottleneck structure, significantly reducing the number of parameters and computational complexity while enhancing the extraction capability of deep features. Furthermore, this invention replaces the original C2f module in the neck network with a novel feature fusion module. This module efficiently fuses multi-scale features, effectively improving the model's ability to represent multi-scale information. Finally, an EMA module is introduced after the neck network and before the detection head to enhance the model's ability to perceive and fuse features at different scales, thereby further improving detection accuracy.

[0034] 2.1 Lightweight Feature Extraction Module; Depthwise separable convolution includes two key operations: depthwise convolution and pointwise convolution. For example... Figure 4 As shown, depthwise convolution uses a single-channel kernel to independently process the spatial dimensions (height and width) of each input channel, effectively extracting local features while maintaining the same number of channels. Pointwise convolution, on the other hand, uses a 1×1 kernel to fuse cross-channel information, freely controlling the number of output feature maps. This decoupled design significantly reduces computational complexity and parameter count while maintaining the same output dimension as standard convolution, achieving efficient feature extraction.

[0035] The activation function is an improved version of the ReLU activation function. For negative input values, it restricts the output to zero or above, and for positive input values, it restricts it to 6 or below, thus producing an output range [0, 6] to prevent the value from being too large. The activation function is specifically designed for mobile devices, improving the stability of quantized models. It is computationally simple and efficient, suitable for lightweight networks. Widely used in the MobileNet series, it balances performance and efficiency. Its formula is shown below: ; In the formula, This is an operation to retrieve the minimum value within the parentheses. This is an operation to retrieve the maximum value within the parentheses. The representative model is the original linear value output by the width and height offset of the bounding box.

[0036] The general inverted bottleneck structure comprises four configurable stages: an initial depthwise convolution to process spatial features, a 1×1 expanding convolution to dynamically increase channel dimensions through expansion ratio, an intermediate depthwise convolution to enhance feature extraction, and a 1×1 projective convolution to compress channels for dimensionality reduction. Through parameterized design, the kernel size for each stage can be customized (supporting 3×3, 5×5, or skipping operations), and flexible downsampling control strategies are supported. Its detailed structure is as follows... Figure 5 As shown, the two optional depthwise convolutions in the optional depthwise separable convolution UIB block have four possible instantiations, resulting in different trade-offs. All convolutional layers employ the ReLU6 activation function and batch normalization, significantly reducing model complexity while maintaining mobile deployment friendliness. This design balances feature representation capability and computational efficiency through a channel expansion mechanism, achieving synergistic optimization of efficient feature extraction and model compression, providing powerful feature representation capabilities for lightweight networks.

[0037] The overall structure of the lightweight feature extraction module, i.e., the feature extraction process, is as follows: Figure 6As shown, the YOLOv8 backbone network is replaced with the following structure. First, the first layer performs initial feature extraction using a standard 3×3 convolution with a stride of 2, BN, and ReLU6. The second layer includes shallow feature refinement using a 3×3 convolution with a stride of 2 and a 1×1 convolution with a stride of 1. The third layer first expands the channels to increase feature representation and then compresses to reduce the number of parameters, including a 3×3 convolution with a stride of 2 and a 1×1 convolution with a stride of 1. The fourth and fifth layers are both UIB structures. The fourth layer mainly contains six specific instances of UIB, specifically configured as follows: additional depthwise separable convolutions (both the initial DW and the middle DW convolutions are 5×5 depthwise convolutions) + four MobileNet inverted bottleneck structures (the middle DW convolution is a 3×3 depthwise convolution) + ConvNext-Like (the initial DW convolution is a 3×3 depthwise convolution). The fifth layer also contains six concrete instances of UIB, specifically configured as follows: additional depthwise separable convolutions (both the initial DW and intermediate DW convolutions are 3×3 depthwise convolutions) + additional depthwise separable convolutions (both the initial DW and intermediate DW convolutions are 5×5 depthwise convolutions) + four MobileNet inverted bottleneck structures (the intermediate DW convolution is a 5×5 depthwise convolution). The structure first extracts features from the input image through multiple convolutions to generate features. Figure 1 and characteristics Figure 2 Then, feature extraction and feature generation are achieved through the UIB structure mentioned above. Figure 3 and characteristics Figure 4 Then, the final features are extracted through multiple convolutional layers. Figure 5 Finally, in the SPPF layer, the input feature maps are pooled in parallel using multiple max-pooling kernels of different sizes (such as 5×5, 9×9, and 13×13) to generate the final feature maps. These feature maps will then be used by the neck network for processing.

[0038] 2.2 Lightweight Feature Fusion Module; This invention enhances the multi-scale feature fusion capability of the neck network by replacing the element-wise summation operation in the C2f module with element-wise multiplication and residual connections. Element-wise multiplication has the ability to map the input to a high-dimensional nonlinear feature space without significantly increasing the network size, achieving excellent performance and low latency with a compact network structure and efficient computation. Based on this, this study constructs a novel feature fusion module, the structure of which is as follows: Figure 7As shown, this module employs a combination of element-wise multiplication and residual connections: through a dual-path feature extraction mechanism, one path's features are activated by ReLU6 and then element-wise multiplied with the other path, achieving dynamic feature selection. This adaptively enhances target-related features and suppresses background noise, particularly improving the signal-to-noise ratio of small targets. The introduction of a 7×7 depthwise convolution further expands the receptive field to 23×23, enhancing the model's ability to perceive target structures and complex textures. Simultaneously, residual connections effectively maintain the integrity of the original features, and DropPath regularization mitigates the risk of overfitting. This module is particularly suitable for challenging scenarios such as small target detection in side-scan sonar images and high-noise backgrounds, significantly improving multi-scale feature representation and generalization capabilities in a lightweight model.

[0039] Element-wise multiplication refers to fusing two input feature vectors through element-wise multiplication (i.e., element-by-element multiplication), thereby achieving high-dimensional nonlinear feature mapping in a low-dimensional input space. The element-wise multiplication formula is shown below. To simplify the analysis of this invention, this invention focuses on scenarios involving single-output channel conversion and single-element input.

[0040] This invention defines , and The element-wise multiplication formula is shown below: ; In the formula, and These are combinations of two different weight matrices and biases, specifically represented as follows: , This is the weight matrix. For bias terms, This is a matrix transpose operation. For the set of real numbers, Input feature vector The corresponding number in The components of each channel, For input vectors The corresponding number in The components of each channel, For the input feature vector, The channel dimension of the input feature. All are channel indexes.

[0041] 2.3 EMA Attention Mechanism; The Efficient Multi-Scale Attention (EMA) mechanism, proposed by Ouyang et al. in 2023, is an innovative lightweight attention architecture. This module significantly reduces computational overhead while preserving complete channel information through channel grouping and dimension reshaping strategies. The overall structure of EMA is as follows: Figure 8As shown, the input feature map is first divided into G sub-features along the channel dimension to ensure a uniform distribution of spatial semantic features. Then, a dual-branch parallel processing is employed, where a 1×1 convolutional branch incorporates a coordinate attention mechanism to extract contextual information across spatial dimensions, while a 3×3 convolutional branch focuses on local detail features. This design, through feature recombination and dual-scale fusion, enables the model to simultaneously capture long-range dependencies and local details, forming rich multi-scale feature representations. Finally, this invention embeds an efficient multi-scale attention mechanism (EMA) before the detection head, allowing the model to retain positional details while integrating multi-scale contextual information, thereby improving the model's ability to perceive and distinguish multi-scale targets in complex underwater scenes.

[0042] Example 2: The underwater complex target detection system based on collaborative optimization of generation and detection provided in this embodiment of the invention includes: The simulation sample generation and screening unit uses CycleGAN to generate side-scan sonar simulation samples in the target domain from source domain images. It uses a pre-trained target detection model to screen the generated simulation samples, retaining samples with confidence scores higher than a preset threshold as high-quality simulation samples. The training dataset construction unit fuses the high-quality simulation samples with the original side-scan sonar dataset to construct an enhanced training dataset. The detection model improvement and training unit constructs an improved object detection model based on YOLOv8 as a baseline, and trains the improved YOLOv8 model using the enhanced training dataset. The target detection inference unit uses the trained improved YOLOv8 model to perform target detection on the input underwater side-scan sonar image.

[0043] Example 3: This embodiment of the invention provides an application of the underwater complex target detection method described above in marine environmental observation, marine engineering development, and maritime emergency search and rescue scenarios.

[0044] This experiment details the process of generating high-quality simulation samples and improving the target detection network, such as... Figure 9 As shown in the figure. The effectiveness of the proposed underwater complex target detection method was then verified and compared with representative methods. The images used in the experiments of this invention were all based on publicly available online datasets (DIOR, Human, SeabedObjects-KLSG, etc.).

[0045] Figure 9This paper demonstrates the overall process of underwater complex target detection proposed in this invention. First, a pre-constructed remote sensing optical dataset is used as the source domain, and the original side-scan sonar dataset is used as the target domain. These are input into a CycleGAN network to generate initial simulation samples. Simultaneously, the original side-scan sonar dataset is used as training input, and Python code is used to train a YOLOv8s target detection model to generate a weight file for the deep learning model, preparing for the subsequent selection of high-quality synthetic samples. Second, inference is performed on the target detection model using the weight file and the initial simulation samples generated by CycleGAN. A confidence threshold of 0.5 is set, and inference results with a confidence threshold greater than 0.5 are considered high-confidence simulation samples. Subsequently, the selected results are manually verified to remove false positives and false negatives, and then traditional data augmentation is performed on the valid samples. Subsequently, based on the original side-scan sonar dataset I, two separate simulation datasets were constructed: an unfiltered original simulation dataset (combining all 283 simulated images of people falling into the water and 991 simulated aircraft images generated by CycleGAN) and a filtered simulation dataset (combining 68 high-confidence simulated images of people falling into the water and 40 simulated aircraft images filtered by the object detection model). These were then used to construct three side-scan sonar datasets that ultimately fused the simulation samples. Finally, the feature extraction and fusion modules of the YOLOv8s model were improved to enhance its multi-scale feature representation capabilities. The improved model was then validated and its performance evaluated based on experimental results. Figure 9 In the technical process of the method proposed in this invention, an improvement is achieved by replacing the original network structure with a lightweight feature extraction and fusion module.

[0046] The practicality and effectiveness of the method proposed in this invention will be verified through specific experiments below.

[0047] The experimental results based on simulated samples are shown in Table 1. The results clearly show that the training accuracy of the YOLOv8s model (key metrics such as mAP@0.5 or mAP@0.5:0.95) significantly improved after adding high-quality simulated datasets. However, low-quality simulated samples severely hampered the improvement in model training accuracy. This result strongly demonstrates that introducing high-quality simulated data has a positive effect on improving the recognition ability and learning performance of the YOLOv8s model. Specifically, the detection accuracy for drowning victims improved by 1.3%, for airplanes by 2.0%, and for boats by 2.0%. The model's mAP@50 improved by 1.7%. Furthermore, the F1 score improved by 3.7%, indicating a comprehensive improvement in the model's ability to characterize classification boundaries, adapt to sample imbalance, and deployment robustness.

[0048] Table 1 Experimental results based on simulation samples

[0049] Table 1 shows that unfiltered simulations include all generated simulation samples, while filtered simulations only include simulation samples filtered by the target detection model.

[0050] Table 2 compares the YOLOv8s baseline model with the improved model, clearly showing significant improvements in model size and detection accuracy. The detection accuracy for drowning victims, aircraft, and shipwrecks improved by 4.4%, 7.9%, and 1.8%, respectively, while the model's mAP@50 improved by 4.7%. The improved accuracy for drowning victims and aircraft indicates that the improved model has enhanced capabilities in small target detection, multi-scale feature extraction, and fusion. Simultaneously, the improved model reduced its parameters by 4.01M, FLOPs by 0.9G, and model size by 7.5MB, validating the effectiveness of the lightweight model approach and demonstrating its ability to meet the deployment and real-time detection needs of edge devices such as underwater robots.

[0051] Table 2. Model size and detection accuracy of YOLOv8s and improved models on the selected simulation dataset.

[0052] also, Figure 10 The changes in mAP@50 during the training process of YOLOv8s and the improved model are shown. When the training reaches 100 times, the mAP@50 of YOLOv8s and the improved model tends to stabilize, and the detection accuracy of the improved model is significantly higher than that of YOLOv8s. Figure 11 The comparison of detection results of the two models on a high-quality simulation dataset is presented. The improved model demonstrates stronger robustness in complex underwater scenarios, effectively mitigating issues such as false positives, false negatives, and duplicate detections present in the baseline model. Furthermore, the improved model assigns significantly higher confidence scores to detected targets, reflecting stronger feature discrimination capabilities and decision reliability. Figure 12 The presentation showcases feature heatmaps and receptive field visualizations, further explaining the performance improvement from a mechanistic perspective. The improved model is more focused on the target region, and its larger receptive field allows the network to capture broader contextual information, particularly enhancing its feature extraction and fusion capabilities for small targets. Simultaneously, the model's ability to suppress interference such as blurred backgrounds and acoustic shadows is significantly enhanced, indicating its ability to learn task-relevant discriminative features more effectively. Comprehensive analysis shows that the structural improvements and training strategies proposed in this invention effectively enhance the model's feature learning ability in underwater side-scan sonar images, demonstrating significant improvements not only in quantitative metrics but also superior generalization and robustness in visual results and feature representation.

[0053] To objectively evaluate the performance advantages of this method in side-scan sonar image target detection, comparative experiments were conducted with current mainstream detection algorithms, including YOLOv8, YOLOv10, YOLOv11, and RT-DETR based on the Transformer architecture. As shown in Table 3, the comparison results cover key indicators such as model parameter count, computational complexity, model size, and detection accuracy.

[0054] Table 3. Comparison of detection accuracy and computational complexity of this invention with other models on the screening simulation dataset.

[0055] Experimental results show that our proposed method achieves a good balance among multiple metrics and significantly outperforms the comparative algorithms: it improves detection accuracy while also achieving model lightweighting. Specifically, although RT-DETR is optimized for real-time detection based on DETR, its model size remains large and computationally complex, failing to effectively balance detection accuracy and efficiency. Compared to YOLOv8s, YOLOv10s and YOLOv11s significantly reduce the number of parameters and computational cost, and enhance multi-scale feature fusion capabilities, but their improvement in small target detection accuracy is limited. In contrast, the proposed method has an extremely small model size and a detection accuracy (mAP@50) of 91.8%, significantly outperforming other algorithms. Although the FLOPs are relatively high, they still meet the requirements for real-time detection. It can be seen that our proposed method successfully achieves an excellent balance between high accuracy and high efficiency in the side-scan sonar image target detection task.

[0056] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for underwater complex target detection based on generation and detection collaborative optimization, characterized in that, The method comprises the following steps: S1, simulation sample generation and screening: using a cycle-consistent generative adversarial network CycleGAN to generate side-scan sonar simulation samples of a target domain from source domain images, screening the generated simulation samples using a pre-trained target detection model, and retaining samples with a confidence higher than a preset threshold as high-quality simulation samples; S2, training data set construction: fusing the high-quality simulation samples with an original side-scan sonar data set to construct an enhanced training data set; S3, detection model improvement and training: constructing a target detection model improved based on YOLOv8 as a baseline, and training the improved YOLOv8 model using the enhanced training data set; The improved YOLOv8 model comprises: a lightweight feature extraction backbone network based on a depth separable convolution and a general inverted bottleneck structure; a lightweight feature fusion neck network using element multiplication and residual connection; and an efficient multi-scale attention mechanism EMA module introduced in front of a detection head; S4, target detection inference: using the trained improved YOLOv8 model to perform target detection on an input underwater side-scan sonar image.

2. The method of claim 1, wherein, In step S1, the source domain images are remote sensing optical images, the target domain is a side-scan sonar image, the pre-trained target detection model is a YOLOv8 model, and the preset confidence threshold is 0.

5.

3. The method of claim 1, wherein, In step S1, the simulation sample screening further comprises a manual verification step to eliminate false positives and false negatives, and data augmentation is performed on the valid samples.

4. The method of claim 1, wherein, In step S3, the lightweight feature extraction backbone network comprises: a depth separable convolution composed of a depth convolution and a point-wise convolution; wherein the depth convolution uses a single-channel convolution kernel to independently process the spatial dimensions of each input channel to extract local features while keeping the number of channels unchanged; the point-wise convolution performs cross-channel information fusion through a 1x1 convolution kernel to freely control the number of output feature maps; a general inverted bottleneck structure containing four configurable stages: a starting depth convolution processing spatial features, a 1x1 expansion convolution dynamically increasing the channel dimension through an expansion ratio, an intermediate depth convolution strengthening feature extraction, and a 1x1 projection convolution compressing the channels to realize dimension reduction; through parameterized design, the convolution kernel size of each stage is customized to support 3x3, 5x5 or skip operations, and flexible down-sampling control strategies are supported; all convolution layers use ReLU6 activation function and batch normalization technology.

5. The method of claim 4, wherein, The feature extraction process of the lightweight feature extraction backbone network is as follows: the input image first goes through a primary feature construction stage to generate multiple groups of basic feature representations; then, these features are integrated and enhanced through a feature interaction and fusion module to produce high-order features with more information and better discriminability; further, a depth refining stage purifies these features; finally, a multi-scale context information aggregation layer adaptively captures and fuses context information of different ranges to output a final feature representation with high information richness, laying a foundation for subsequent decoding and prediction tasks. The activation function is an improved version of the ReLU activation function, which limits the output to zero or above for negative input values and to 6 or below for positive input values, resulting in an output range [0, 6], expressed as: ； wherein is a minimum operation on the values in the parentheses, is a maximum operation on the values in the parentheses, represents the original linear values output by the model for the width and height offsets of the bounding box.

6. The method of claim 1, wherein, In step S3, the lightweight feature fusion neck network adopts the combination of element multiplication and residual connection, and through a double-path feature extraction mechanism, the features of one path are multiplied with the features of another path after ReLU6 activation, realizing dynamic feature selection to adaptively enhance target-related features and suppress background noise; by introducing a specific depth convolution, the model receptive field is expanded to enhance the model's perception of target structure and complex texture; The element multiplication operation refers to fusing two input feature vectors through element-level multiplication, thereby realizing high-dimensional nonlinear feature mapping in a low-dimensional input space; taking a single output channel conversion and a single element input as an example, definitions of , and are as follows: ； In the formula, and These are combinations of two different weight matrices and biases, specifically represented as follows: , This is the weight matrix. For bias terms, This is a matrix transpose operation. For the set of real numbers, Input feature vector The corresponding number in The components of each channel, For input vectors The corresponding number in The components of each channel, For the input feature vector, The channel dimension of the input feature. All are channel indexes.

7. The method of claim 1, wherein, In step S3, the efficient multi-scale attention mechanism EMA module processes the characteristics of the underwater side-scan sonar image. Through channel grouping and dimension reshaping strategies, the input feature map is divided into G sub-features in the channel dimension to adapt to the complex distribution of target multi-scale and background clutter in the sonar image. On this basis, a double-branch parallel processing structure is adopted: the 1×1 convolution branch incorporates a coordinate attention mechanism to effectively capture the spatial context correlation along the track and perpendicular to the track direction in the underwater side-scan sonar image, enhancing the model's perception of fuzzy and low-contrast acoustic targets. The 3×3 convolution branch focuses on the enhancement and extraction of local details and texture features in the sonar image to suppress acoustic noise interference, thereby improving the detection accuracy and robustness in underwater complex acoustic environments.

8. The method of claim 1, wherein, In step S4, the trained improved YOLOv8 model is used to detect multi-scale targets in underwater side-scan sonar images, including swimmers, aircraft, and ships.

9. An underwater complex target detection system based on generation and detection co-optimization, characterized in that, The system is used to implement the underwater complex target detection method based on generation and detection collaborative optimization as claimed in any one of claims 1-8, comprising: A simulation sample generation and screening unit uses the CycleGAN to generate side-scan sonar simulation samples from source domain images, and uses a pre-trained target detection model to screen the generated simulation samples, retaining samples with a confidence higher than a preset threshold as high-quality simulation samples; A training data set construction unit fuses the high-quality simulation samples with original side-scan sonar data sets to construct an enhanced training data set; A detection model improvement and training unit constructs a target detection model improved based on YOLOv8, and trains the improved YOLOv8 model using the enhanced training data set; A target detection inference unit uses the trained improved YOLOv8 model to perform target detection on input underwater side-scan sonar images.

10. The underwater complex target detection method as claimed in any one of claims 1-8 is applied in marine environment observation, marine engineering development, and marine emergency rescue scenarios.

Images