Attention-Fused Shallow Network for Underwater Object Detection

Although generic object-detection methods have achieved a great deal, it needs more exploration for underwater images. Underwater object detection (UOD) is associated with challenges such as degraded image quality, low visibility, low contrast, colour shift, and limited computational capacity availability on the deployment environment. Moreover, most previous studies on deep learning-based underwater object detection have generally concentrated on increasing detection accuracy by utilizing huge networks. This work proposes a two-stage method where in the first stage, underwater images are enhanced based on PCA-fusion method. This step involves mutiple image enhancement steps such as color correction process, followed by the White Patch Retinex Algorithm for white balancing. On the colour-corrected image, we apply global histogram equalization, unsharp masking, and median smoothing separately to improve the contrast, sharpen the image, and reduce the white patch noise. Then, we create a single enhanced image by combining the results of the three methods using the principal component analysis (PCA) based fusion method. Finally, in the second stage, the enhanced image is passed as an input to the attention-fused lightweight single-stage object detection model for localization and classification. Experimental results show that the proposed method outperforms the state-of-the-art algorithms based on various image enhancement and object detection evaluation metrics on the URPC2019 dataset.