DPCDet: Dual-path downsampling and cross-attention enhanced detector for forward sonar image

Forward-looking sonar is crucial for underwater perception, yet its images suffer from low resolution, weak texture, and low signal-to-noise ratio, challenging general detection models. Existing detection networks lose details through successive down-sampling, further weakening limited semantic clues; meanwhile, residual-based cross-layer fusion lacks dynamic modeling of feature importance across stages. In addition, sonar noise can degrade image quality and adversely affect detection accuracy. To address these issues, we propose a lightweight sonar image detector named DPCDet, featuring two task-oriented architectural modules. Dual-Path Downsampling Fusion (DPDF) module partitions the input channels and combines a learnable strided-convolution path with a statistical path that jointly preserves local mean and peak responses. Cross-Stage Attentive Fusion (CSAF) module derives spatially varying source weights from consecutive same-resolution stages and their joint statistics, adaptively balancing local details and contextual information within each block. Furthermore, denoising is introduced as an optional preprocessing step to enhance input image quality and mitigate noise interference. Experiments on the self-built NKMESD dataset and the public FLSMDD sonar dataset demonstrate that the proposed method achieves state-of-the-art performance in rotated object detection tasks while optimizing model parameter count and computational efficiency. This work provides a technical solution for robust target detection in complex underwater environments.