Underwater target detection with an attention mechanism and improved scale

The light problem and the complicated environment of underwater images make target detection difficult. These images are usually blurry because tiny inorganic and organic particles in the water have a great impact on light. To solve this problem, we add squeeze and excitation modules after the deep convolution layers of the YOLOv3 model to learn the relationship between channels and enhance the semantic information of deep features. In addition, many small targets will lose too much information after five downsamples. This is not conducive to detection. By expanding the detection scale, we combine the deep semantic information with the location information of the shallower layer to improve the detection performance of small targets. The experimental results show that the YOLOv3-brackish model greatly improved the detection of small fish, crabs, shrimp and starfish. In addition, there were minor improvements in the detection of big fish and jellyfish. The mean average precision increased by 4.43%.