Lightweight underwater object detection based on image enhancement and multi-attention

Underwater object detection, which is crucial to the exploration and exploitation of marine resources, remains a challenge because noisy, weak contrast, and color distorted images are provided as sources of supervision. To address the issues of low detection accuracy caused by imprecise images, and inefficiency due to huge amount of parameters in most deep neural networks, this paper proposed a novel lightweight deep learning model with image enhancement and multi-attention. First, image enchancement algorithm MSRCR is applied to enhance image quality in order to improve the training effect of deep learning model. Then, YOLOX is used as baseline model and GhostNet is utilized as backbone network in order to reduce computation budget. Finally, a multi-attention module LCR considering level, channel and spatial domains is divised and integrated into the feature pyramid network to enhance feature learning ability and detection accuracy. Experimental result shows that on the considered datasets our model achieves an mAP of 77.32 \(\%\) $\%$ and a size of 18.5MB, 1.25 \(\%\) $\%$ higher and 46.4 \(\%\) $\%$ less than the values of baseline network, indicating that our method achieves a superior detection precison while keeping model lighweight.