In this study, style transformation based on CycleGAN was performed to generate SSS images, utilizing datasets from two domains: real-world SSS images and synthetic images for style transformation. The SeabedObjects-KLSG dataset, published by Huo et al., is a publicly available collection of 1190 underwater SSS images gathered over 10 years [15]. These images are categorized into five groups—shipwrecks, aircraft, drowning victims, mines, and seafloor—to support automatic underwater object detection research. For this study, 402 images from three categories—334 shipwrecks, 34 aircraft, and 34 drowning victims—were selected, while other categories such as mines and seafloor were excluded. Mines were excluded because their specifications vary widely across different types, and the shadows they cast are generally small, making them less relevant to the objectives of this study. The seafloor category was also omitted, since the focus of this study is on object-level representation rather than background modeling. Fig. 1 presents sample images from the SeabedObjects-KLSG dataset, showcasing real-world examples of SSS imagery.

Beyond the SeabedObjects-KLSG dataset, the availability of public SSS datasets containing object instances is extremely limited. Due to security restrictions and data sensitivity, most existing studies have relied on private datasets, which makes reproducibility difficult. Moreover, many works have primarily focused on mine or mine-like objects [16–18], while comparatively little attention has been given to objects with more complex structural features. The SeabedObjects-KLSG dataset remains one of the few publicly accessible resources and has therefore been widely adopted in related research [19,20]. Other available datasets mainly contain categories unrelated to this study, including glaciers and walls [21], pipes, mounds, and platforms [22], walls [23], pipelines [24], and seagrass, rocks, and sand [25]. Consequently, the SeabedObjects-KLSG dataset was selected as the most suitable basis for the experiments in this work.

To complement real-world data, a rendering image dataset was created using Blender, a 3D computer graphics software. This dataset includes 3D models of ships, aircraft, and human figures, transformed through rotation, deformation, and other adjustments. Camera angles were varied to capture different perspectives, while light sources were altered to generate diverse shadow effects. The rendering dataset comprises 117 shipwreck images, 104 aircraft images, and 100 drowning victim images. Fig. 2 illustrates examples from this dataset, demonstrating the variety achieved through controlled 3D rendering.

This study implements a style transformation model based on CycleGAN to generate SSS images from rendered images of various 3D models. A limitation of the traditional CycleGAN structure is that, with a small amount of data, the generated images may not sufficiently reflect the characteristics of real SSS images. To address this issue, we introduced a multi-resolution training structure and a shadow area loss function. The multi-resolution training method enables the model to simultaneously learn from images of varying resolutions, allowing the network to learn diverse features effectively. The shadow area loss function compares the shadow regions extracted from the original and generated images, training the network to preserve the original object’s shadow shape while accurately reflecting the shadow characteristics of SSS images. The network structure used for generating SSS images based on CycleGAN in this study is presented in Fig. 3.

CycleGAN is a GAN-based model designed to learn image translation between two different domains. It is composed of two generators and two discriminators. These two neural networks, the generators, and discriminators, interact with each other, progressively learning to generate and discriminate realistic images. A key advantage of CycleGAN is its ability to learn image translation between two domains through unsupervised learning without requiring a one-to-one matched dataset. Additionally, cycle consistency helps minimize distortions and information loss during the translation process, enabling more consistent and reliable image transformations [26]. Moreover, the lack of paired ground-truth supervision does not guarantee the preservation of high-frequency structures (e.g., edges and shadow boundaries). Adversarial loss primarily enforces distribution-level realism, while cycle-consistency loss emphasizes reversibility of content rather than exact edge fidelity; as a result, generated images may not consistently retain fine details.

Fig. 3 illustrates the training structure for the domain s (SSS images) in this study. The generator GRS transforms images from domain r (rendered images) into images in domain s, while the generator GSR converts images from domain s back into images in domain r. The discriminator DS determines whether the generated images belong to the real domain s, and the discriminator DR determines whether the generated images belong to the real domain r. Unlike the conventional CycleGAN, the proposed model incorporates additional components tailored to SSS imagery: a k-means–based Shadow Extractor that highlights shadow regions, and a Shadow Region Loss that enforces the preservation of these regions during translation. Furthermore, multi-resolution training enables the model to capture both global structure and fine details more effectively. These modifications allow the framework not only to perform domain translation but also to retain the critical high-frequency cues, such as edges and shadow boundaries, which are essential for reliable analysis of SSS data.

Due to the limited number of available SSS images, the conventional CycleGAN structure faced limitations in generating SSS images. To address this limitation, this study proposes a multi-resolution training structure. The multi-resolution learning structure refers to a structure in which the generator and discriminator are trained at multiple resolutions for a single image, allowing for more diverse and efficient learning [27]. This study conducts training at four resolutions: 256 × 256, 128 × 128, 64 × 64, and 32 × 32. Fig. 4 provides a detailed overview of the generator and discriminator structure used for multi-resolution training.

The generator is based on the U-net structure employed in the traditional CycleGAN generator, with an added feature that generates images at the four specified resolutions from each decoder layer. The four generated images at different resolutions are passed to the discriminator, which extracts features from each resolution and then combines them to produce a decision value. The real images are also converted into four resolutions to match the discriminator’s input format. The output of the discriminator adopts a PatchGAN structure, where the image is divided into small patches, and each patch is evaluated, resulting in an output of (1, 30, 30) shape. This enables the model to learn more detailed and localized information across the image [28].

In this study, four loss functions—adversarial loss, cycle consistency loss, identity loss, and shadow area loss—are used to train the model.

This study conducted experiments using both the SSS and rendered image datasets. The SSS image dataset consisted of 403 images, including 334 shipwrecks, 34 aircraft, and 35 drowning victims. The Blender-generated image dataset comprised 321 images, 117 shipwrecks, 104 aircraft, and 100 drowning victims.

The model was trained for 1000 epochs, and data augmentation techniques such as CenterCrop, HorizontalFlip, and VerticalFlip were applied. CenterCrop is a technique that crops the central part of the image to adjust its size, while HorizontalFlip and VerticalFlip flip the image horizontally and vertically, respectively. The input data was normalized with a mean of 0.5 and a standard deviation of 0.5. The learning rate was set to 0.0001, and optimization was performed using Adam with β = (0.5, 0.999), consistent with the settings reported in the original CycleGAN study.

Weighted loss functions were applied with the following values: λ1=10, λ2=0.4, and λ3=0.6. The relatively large weight of λ₁ enforced the preservation of structural information during domain translation, thereby mitigating mode collapse as well as geometric and topological distortions. The shadow weight λ₃ emphasized the retention of shadow regions, a critical feature of sonar images, whereas λ₂ prevented unnecessary modifications to samples that already belong to the target domain. This configuration preserved the structural integrity of the original images while maintaining shadow fidelity and suppressing undesired transformations. The weights were empirically determined through repeated experimentation and refined based on expert feedback from sonar specialists to achieve results that closely resemble real sonar images.

To analyze the impact of each loss weight (λ), we conducted experiments in which each λ value was individually minimized to 0.1. In addition, experiments without multi-resolution training were performed, with the results presented in Fig. 5. When λ1 (cycle-consistency) was minimized, the generated images exhibited severe structural distortions of the objects, demonstrating its importance in preserving overall geometry. When λ2 (identity) was minimized, some regions showed blurred boundaries, highlighting its role in constraining unnecessary changes to samples from the target domain. When λ3 (shadow) was minimized, the generated outputs exhibited inaccurate shadow orientations and partial degradation of shadow regions, underscoring the necessity of shadow loss for realistic sonar imagery. Finally, disabling multi-resolution training produced images lacking the distinctive characteristics of sonar data, making them clearly distinguishable from real sonar images.

The generated images exhibit the features of SSS imagery, as presented in Fig. 6. The highlights are caused by the direct reflection of sound waves off objects, and the shadows formed in areas where objects obstruct sound waves are prominently displayed. The overall noise in the images, resulting from various sources of underwater interference, is also effectively represented. These features demonstrate that the model effectively captures the distinctive properties of SSS imagery.

To evaluate the generated SSS images against other generative models, we trained existing CycleGAN models—Unet128, Unet256, Resnet-06, and Resnet-09—and compared their generated outputs with those from the proposed model. When analyzing the outputs of these existing models, we observed that traditional models struggled to preserve object structure, with significant distortions occurring in the background and shadows, resulting in unsuccessful image generation. In some cases, the images lacked typical SSS characteristics, presenting only irregular patterns in place of the original features. In contrast, the model proposed in this study consistently preserved the shape of the objects, background, and shadow while also effectively capturing key characteristics of SSS images, such as the highlights of object surfaces, shadow details, and noise (Fig. 7).

To evaluate the computational requirements of the proposed model, we compared the number of parameters, multiply-accumulate operations (MACs) and floating-point operations (FLOPs) with those of the baseline CycleGAN. Table 1 summarizes the results for an input size of 1 × 256 × 256. The proposed model contains 15.45 M parameters, which is comparable to the 14.12 M parameters of CycleGAN. The total MACs decreased from 59.16 to 17.06 G, and the total FLOPs decreased from 118.32 to 34.11 G, representing reductions of approximately 71.2% in both metrics.

This result is attributed to structural design differences. Conventional CycleGAN applies nine residual blocks on relatively high-resolution feature maps, which considerably increases computational cost despite having a similar number of parameters. Because MACs/FLOPs scale with the spatial dimensions of feature maps and the number of convolutional operations, repeated residual blocks substantially amplify the overall computational load in CycleGAN. In contrast, the proposed model distributes computation across multiple resolutions, executes a substantial portion at lower resolutions, and uses an optimized discriminator that avoids redundant operations. As a result, the proposed framework maintains a similar parameter scale while reducing MACs and FLOPs by more than two-thirds.

In this study, we used the YOLOv5 model [30] to evaluate the performance of ATR when trained with SSS images generated by different models, including Unet128, Unet256, ResNet-06, ResNet-09, and the proposed CycleGAN-based method. The experiment was conducted over 200 epochs, using real SSS data split into training and testing sets in a 50:50 ratio. The YOLOv5n model, known for its lightweight structure and fast inference speed, was trained from scratch without pre-trained weights. Following initial training, we augmented the training dataset by adding 117 shipwreck images, 104 aircraft images, and 100 drowning victim images generated by the model described above. A comparative analysis was conducted to evaluate the effectiveness of these generated datasets in improving ATR performance.

Table 2 summarizes the quantitative evaluation results, including object detection performance evaluated with YOLOv5 and image quality assessed using Fréchet inception distance (FID). The proposed method yielded the highest mAP for shipwreck (0.898) and drowning victim (0.942) detection, demonstrating superior performance in these categories compared with other models. For aircraft detection, ResNet-06 achieved the highest mAP (0.679), while the proposed method obtained a lower mAP of 0.508. To further evaluate image quality, we employed FID as an objective image similarity metric. Since the conventional InceptionV3 model used for FID is trained on natural images and thus unsuitable for sonar imagery evaluation, we replaced it with a ResNet18 model fine-tuned on sonar images. The FID results obtained using this approach are summarized in Table 2. The experimental results show that our method achieved the lowest FID scores for the Ship and Victim classes, while ResNet-06 yielded the lowest FID for the Plane class. Upon further analysis, we observed that the real sonar Plane images used in training contained numerous complex and cluttered signals. These signals likely resulted from structural damage during crashes or from prolonged underwater exposure, which caused corrosion and deformation of the fuselage and wings, producing intricate internal structural patterns. In contrast, the images generated by our method were derived from 3D models without surface damage or corrosion, leading to cleaner representations where the acoustic responses of the fuselage and wings were expressed with reduced noise. This seems to imply that the ResNet-based model, which was trained to emphasize irregular and noisy features, assigned relatively lower FID values to real Plane images compared with our generated counterparts.

To evaluate the impact of the shadow area loss function on training, the proposed model was trained with the weight of the shadow area loss function set to 0, and the generated results were compared.

The experimental results, shown in Fig. 8, reveal that abnormal shadow patterns appeared in the background in the model without the shadow area loss function. Additionally, the shapes of the object shadows were unstable. These findings indicate that the shadow area loss function plays a crucial role in preserving the shadow shapes of objects in the input images and ensuring that the overall shadow characteristics of SSS images are represented consistently and accurately.