23 October 2023 in IEEE Transactions on Geoscience and Remote Sensing
Super-Resolution is a technique for recovering image details based on available information, avoiding image quality degradation by increasing an image’s resolution. Despite the great success of deep learning-based SR models, they are rarely applied to underwater sonar scenarios due to the lack of underwater sonar datasets and the difficulty in recovering texture details. To address these challenges, we propose a multi-headed generative adversarial network (MHGAN) for the super-resolution of underwater sonar images. The main body of the generator is designed with the introduced Residual Simple-dense Self-calibrated Block (RSSB), which combines Self-calibrated Attention and Pixel Attention to improve generalization capability. Additionally, the discriminator employs a novel multi-headed U-net structure to acquire detailed texture information at multiple patch sizes of different scales, along with the generated image and High-Resolution(HR) image details. Moreover, we introduce the correction loss, which acts as a pixel-level correction to help improve the quality of the output images, and multi-perceptual loss based on VGG19 to extract multi-scale features for comparison. Meanwhile, we also introduce the Underwater Sonar Dataset for Super-Resolution (USDSR). We perform the visual and quantitative comparison, ablation experiments, and model analysis with state-of-the-art methods on both the public KLSG dataset and the built USDSR dataset. Experimental results demonstrate the effectiveness and superiority of our proposed MHGAN model.
17 March 2025 in IEEE Transactions on Instrumentation and Measurement
Object detection in sonar images is challenging since the performance based on generic object detectors is generally not well on sonar datasets and the prior for the scale distribution of the objects is usually ignored. To overcome these issues, in this article, multi-scale graph channel attention (MGCA) detectors are proposed, in which Multi-Scale Box Copy-Paste is designed for data augmentation and Graph Transformer Channel Attention (GTCA) based on ConvNeXt is introduced to learn powerful visual representations for sonar images. GTCA contains three basic modules: graph structure transformation module for feature maps, graph convolution module for aggregating and updating information and feed forward network module for feature transformation. With Multi-Scale Box Copy-Paste to balance scale distribution and channel attention network in a graph form to recalibrate feature maps, MGCA detectors can effectively detect multi-scale objects in sonar images through transfer learning. The MGCA detector with Cascade R-CNN detection head achieves detection accuracies of 95.6 mAP and 72.7 mmAP on sonar common target detection dataset (SCTD), outperforming the previous best detector with increase of 2.8 mAP and 11.8 mmAP. Extensive experimental results and ablation studies on two sonar datasets demonstrate the superiority of the proposed method compared with state-of-the-art methods.