To enhance both efficiency and accuracy in detection, this study employs the k-means++ (Arthur and Vassilvitskii, 2007) technique to supplant the k-means approach, initially employed in YOLOv7, for clustering anchor boxes within the dataset. In the conventional k-means method, the first phase involves the random generation of n cluster centers from the data samples. Subsequently, the Euclidean distance between each sample and the cluster centers is computed, and the sample is assigned to the cluster center exhibiting the smallest Euclidean distance. In the subsequent phase, the cluster centers are reevaluated, and samples are reclassified. This iterative process is repeated until the cluster centers reach stability.

The k-means++ method represents an enhancement over the conventional k-means approach. Unlike generating all cluster centers randomly in a single instance, k-means++ generates one cluster center at a time. It calculates the Euclidean distance D(x) between all samples and the cluster center, subsequently deriving the probability of each sample being chosen as the next cluster center through the Equation 1.

Subsequently, the next cluster center is chosen via the roulette wheel selection method. This sequence of steps is reiterated until n cluster centers are generated. After this stage, the ensuing process resembles that of the conventional k-means algorithm: the cluster centers are updated, samples are reclassified, and these steps are iterated until the cluster centers achieve stability. While the k-means++ algorithm invests more time in selecting initial cluster centers, once these initial centers are established, the convergence speed accelerates, yielding cluster centers that hold greater representativeness. This approach mitigates the challenge of becoming trapped in local optima.

In current neural networks, the majority typically employ static convolutional kernels. However, recent research on dynamic convolutions suggests calculating relevant weights based on the input and linearly combining n convolutional kernels according to these weights. This makes the convolution operation dependent on the input, leading to a significant improvement in neural network accuracy. The experimental results from (Li et al., 2022) demonstrate that the use of ODConv enhances the detection performance for small targets. Therefore, in this study, all convolutional operations in the YOLOv7 backbone network are replaced with ODConv to enhance the detection performance of the network.

The core innovation of ODConv lies in its multi-dimensional dynamic attention mechanism. Traditional dynamic convolution typically achieves dynamism only in the dimension of the number of convolutional kernels, by weighting and combining multiple kernels to adapt to different input features. ODConv extends this concept further by dynamically adjusting not only the number of convolutional kernels but also three other dimensions: spatial size, input channel number, and output channel number. This means that ODConv can adapt more finely to the features of input data, thereby improving the effectiveness of feature extraction.

Additionally, ODConv employs a parallel strategy to simultaneously learn attention across different dimensions. This strategy allows the network to efficiently process features on each dimension while ensuring complementarity and synergy among the dimensions. This is particularly beneficial for handling complex features in SSS images. The network structure is illustrated in Figure 2 .

The output after ODConv can be expressed using the Equation 2.

where a represents the attention parameter for the spatial dimensions of the convolutional kernel, b represents the attention parameter for the input channel dimensions, c represents the attention parameter for the output channel dimensions, d represents the attention parameter for the convolutional kernel W.

The incorporation of attention mechanisms within neural networks draws inspiration from human visual attention, enhancing feature extraction by assigning distinct weights to various channels within neural network feature layers. This strategy enables the model to concentrate on pertinent information while disregarding irrelevant data, leading to resource conservation and augmented model performance. Several mainstream attention mechanisms, such as SE-Net (Hu et al., 2018), ECA-Net (Wang et al., 2020), BAM (Park et al., 2018), CBAM (Woo et al., 2018) and GAM (Liu et al., 2021), have been demonstrated to enhance the detection performance of models.

The GAM represents a form of global attention mechanism that curtails information loss and amplifies interactions across global dimensions. Consequently, the neural network’s aptitude for extracting target features is bolstered. The schematic depiction of the GAM module structure is presented in Figure 3 .

GAM employs a sequential channel-spatial attention mechanism with the aim of amplifying global inter-feature interactions while reducing information dispersion. In the channel attention submodule of GAM, a three-dimensional configuration is employed to preserve tridimensional information. The input feature map undergoes dimensional transformation and subsequently undergoes an MLP operation. The result is then reverted to the original dimension, and a sigmoid function is applied to produce the final output.

In the spatial attention submodule, aimed at intensifying focus on spatial information, two convolutional layers facilitate spatial data fusion. Initially, a convolution employing a kernel size of 7 is executed to diminish channel count and computational complexity. Subsequently, another convolution with a kernel size of 7 enhances the number of channels while maintaining uniform channel consistency. The resulting output is then processed through a sigmoid function.

In order to enhance the detection performance of the detection network, we introduced GAM modules at five distinct locations in the neck network. The architecture of the YOLOv7 network with added GAM modules, as well as the specific structures of individual sub-modules within the network, are illustrated in Figure 4 .

The bounding box regression function holds a pivotal role in object detection by enhancing object localization accuracy, accommodating objects of varying scales, rectifying object orientations and shapes, and bolstering algorithmic robustness. This collective functionality contributes significantly to the advancement of object detection algorithms.

However, the majority of current research on Intersection over Union (IoU) (Yu et al., 2016) assumes that the training data consists of high-quality samples, with their primary focus being on enhancing the fitting capability of bounding box regression loss functions, such as Generalized-IoU (GIoU) (Rezatofighi et al., 2019), Distance-IoU (DIoU) (Zheng et al., 2020), Complete-IoU (CIoU) (Zheng et al., 2020), and Efficient-IoU (EIoU) (Zhang et al., 2022b), as shown in Table 1 , where their advantages and disadvantages are compared. Yet, when dealing with datasets that contain a significant amount of inaccurately annotated low-quality data, blindly intensifying the fitting ability of the bounding box regression loss function can have detrimental effects on the model’s learning process.

In SSS imagery, targets are highly susceptible to noise interference in the generated images, which presents significant challenges for annotation. In the process of manual annotation, inaccuracies inevitably arise, as illustrated in Figure 5 . If the annotation boxes are initially flawed, when an excellent detection model generates high-quality anchor boxes for low-quality sample data, the loss function L_IoU will have a relatively large value, leading to a substantial gradient gain. In such cases, the model will learn in an unfavorable direction. This phenomenon is particularly relevant in the context of scientific research and analysis for SSS imagery.

To address the issue of poor quality in underwater SSS data, we introduce the WIoU (Tong et al., 2023) as the bounding box loss function. This aims to alleviate the impact of low-quality anchor boxes generated during annotation. The WIoU function employs a dynamic non-monotonic focus mechanism that evaluates anchor box quality through outliers, instead of IoU. This approach furnishes a judicious gradient allocation strategy, curbing the competitiveness of high-quality anchor boxes while attenuating detrimental gradients arising from low-quality instances. Consequently, WIoU prioritizes anchor boxes of moderate quality, ameliorating detector performance overall.

The symbols defined in WIoU are illustrated as shown in Figure 6 . In this figure, the blue box represents the smallest bounding box, and the red line represents the line connecting the centers of the true box and the predicted box, where the union area is denoted as S u = w h + w g t h g t − W i H i .

The WIoU methodology, founded on distance metrics, incorporates a two-tier attention mechanism known as WIoU v1. WIoU v1 can be represented by Equations 3, 4.

where ℒ I o U ∈   [ 0 , 1 ] , W_g and H_g are the dimensions of the minimum bounding box, ( x , y ) and ( x g t , y g t ) represent the center coordinates of the predicted box and the ground truth box.

Subsequently, building upon WIoU v1, the incorporation of outliers is achieved through the Equation 5.

Finally, a non-monotonic focus coefficient β is formulated and integrated into WIoU v1. As a result, we obtain Equation 6.

A reduced outlier score implies a higher quality anchor box, yielding a diminished gradient gain assigned to it. Consequently, the bounding box regression concentrates on anchor boxes of intermediate quality. In contrast, anchor boxes exhibiting larger outlier scores are allocated lesser gradient gains, effectively curtailing the generation of significant harmful gradients from low-quality instances. Notably, as ℒ I o U ¯ remains dynamic, the categorization threshold for anchor boxes’ quality also remains adaptive. This adaptability empowers WIoU to judiciously allocate gradient gains that are suitable for real-time scenarios, enhancing its effectiveness in each instance.

The experiments presented in this study were carried out on an Ubuntu 20.04 system, serving to corroborate the efficacy of the proposed enhanced detection algorithm. Detailed configuration parameters of the system are provided in Table 2 .

When evaluating the detection performance of the improved YOLOv7, we employed evaluation metrics including Recall (R), Precision (P), Average Precision (AP), and mean Average Precision (mAP). The calculation methods of these four indicators can be expressed by Equations 7–10 respectively.

Within the array of evaluation metrics mentioned, True Positive (TP) signifies the tally of correctly identified positive samples, False Positive (FP) corresponds to the count of erroneously identified negative samples, and False Negative (FN) stands for the tally of positively labeled samples that remain undetected. The variable N represents the overall number of detected categories.

In the data acquisition phase, we deployed objects of two types, namely cylindrical and conical structures, as detection targets in the experimental marine area. We utilized the SS3060 dual-frequency SSS as the detector for data collection. The SSS and detection targets are illustrated in Figure 7 . The size of the SSS is 100mm in diameter and 1250mm in length, with a weight of 25kg in air and 12kg in water. And the performance parameters of SSS are presented in Table 3 . For the experiment’s execution, the SSS was affixed beneath an unmanned boat. The utilization of GPS signals emanating from the unmanned boat enabled the verification of congruence between features visible in the SSS images and the physically predetermined targets. This methodology thereby facilitated the creation of a dataset characterized by high quality.

After deploying the targets, to ensure the diversity of the collected dataset, we employed two different survey paths in the target water area to perform a comprehensive scan of underwater targets. The placement of the targets and the scanning paths are illustrated in Figure 8 . In the figure, the lateral distance between the targets is approximately 50 meters, and the longitudinal distance is approximately 100 meters. Due to the influence of underwater currents, some degree of deviation in this distance is inevitably present.

Due to the complex and variable underwater environment, as well as the susceptibility of images to noise interference, the images acquired using SSS also exhibit significant variations, as shown in Figure 9 .

Discerning distinct target features within SSS images presents a formidable challenge. Manual annotation subsequent to data collection is arduous, making on-site, real-time labeling the optimal strategy. To attain the real-time processing of SSS images, we adopt a tactic wherein image segments are extracted from the sonar waterfall plot at intervals of 30 seconds, illustrated in Figure 10 . This approach facilitates the annotation of targets on SSS images in real-time, while accounting for the field environment and GPS coordinates.

Furthermore, the targets occupy a minuscule proportion within the complete SSS image. Training the network directly with large-scale SSS images would generate an excessive number of negative samples, potentially impeding the training process and squandering computational resources. Moreover, considering practical applications, the network needs to be deployed on resource-constrained underwater autonomous vehicles, making it imperative to restrict the image size fed into the detection network. To address these challenges, we partitioned the images into diminutive patches with dimensions of 200 × 200. Each patch features a 50-pixel overlap to prevent the loss of target characteristics. From these patches, we selectively identified those containing targets for training, significantly reducing the generation of irrelevant negative samples stemming from extraneous background information. Similarly, during the detection phase, we performed the same cropping operation before inputting the complete image into the detection network.

Finally, we filtered out unusable data and conducted data augmentation using high-quality data, yielding a total of 975 sample images. These images include 293 Cones, 318 Cylinders, and 364 Non-target instances. (“Non-target” refers to miscellaneous items on the seafloor, such as rocks or accidentally dropped artificial objects, which were not intentionally deployed by us. Despite not being the primary focus of the experiment, these Non-target items share certain similarities with the intentionally deployed targets. Including them in the dataset is essential, as their presence could potentially impact our ability to detect the deployed targets.) These samples were then randomly divided into training, validation, and test sets in a 7:1:2 ratio, with the specific number of samples for each set as shown in Table 4 .

To validate the effectiveness of the algorithm proposed in this study for detecting small targets in SSS imagery, we tested the algorithm on a real dataset collected during our sea trials. The variations in various loss functions and accuracy metrics during the training process are illustrated in Figure 11 .

To ensure that all the introduced modifications exerted a positive influence on the network, a sequence of ablation experiments was carried out. The results of these experiments are presented in Table 5 . In the table, mAP@0.5 represents the average precision at an IoU threshold of 0.5, while mAP@0.5: 0.95 represents the average of mAP values at IoU thresholds ranging from 0.5 to 0.95. It is apparent that the integration of k-means++, ODConv, GAM, and WIoU enhancements has resulted in an improved detection performance of the original YOLOv7 model on our assembled SSS dataset. The comparison of Precision-Recall (PR) curves on the test set between the improved YOLOv7 network and the original YOLOv7 network is shown in Figure 12 , while the comparison of confusion matrices is shown in Figure 13 . From Figure 12 , it can be observed that the improved YOLOv7 network achieved an average precision improvement of 5.05% on the test set.

From Figures 12 , 13 , it can be observed that the improved YOLOv7 network demonstrates a noticeable enhancement in the detection performance of Non-target objects. In Figure 12 , the PR curve of the enhanced YOLOv7 network for the Non-target category shows a value of 0.909, which represents an improvement of 0.095 compared to the original network’s 0.814. In Figure 13 , within the improved YOLOv7’s confusion matrix, the Non-target category registers a value of 0.94, as opposed to the original network’s 0.92, marking a 0.02 improvement.

In addition, our experimental results provide evidence of the enhanced network’s superior performance in detecting Non-target objects, as depicted in Figure 14 . The original YOLOv7 network misclassified Non-target objects as Cylinder and Cone, whereas the improved YOLOv7 network can accurately identify Non-target categories. This advancement has reduced the false detection rate for Non-target, which holds significant practical significance in engineering applications. During the search process, it prevents wasting time on Non-target objects.

Furthermore, a comparative analysis was conducted between our enhanced detection algorithm and prominent detection networks to validate the efficacy of the proposed methodology. The comparative visualization of detection outcomes is illustrated in Figure 15 . Detailed detection metrics are presented in Table 6 . These findings collectively furnish compelling evidence for the superior performance of the approach proposed in this paper within the domain of small target detection using SSS.

The results illustrated in Figure 15 provide empirical validation of the efficacy of the approach introduced in this research. As demonstrated in columns (2), (3), and (4) of Figure 15 , some mainstream detection networks often exhibit mis-detections when accurately distinguishing between the categories of cylindrical and conical objects. In contrast, the proposed method in this paper demonstrates accurate detection for objects that are challenging to differentiate, with higher probability values assigned. This highlights the superiority of the algorithm presented in this paper.

Nonetheless, it is important to note that the enhanced network in this study does exhibit certain limitations. For example, as depicted in column (6) of Figure 15 , all networks misclassify a Non-target as a Cone. This misclassification arises due to the distinct shadow surrounding the Non-target and the similarity in the size of the bright spot to the Cone category, resulting in a false positive detection. At present, there is a lack of definitive solutions for scenarios in which acoustic image features exhibit extremely high similarity, yet the actual objects belong to different categories. Using a higher-precision device to acquire images with increased resolution may be beneficial for addressing this issue.