MCR-YOLO model for underwater target detection based on multi-color spatial features

Within the fields of underwater robotics and ocean information processing, computer vision-based underwater target detection is an important area of research. Underwater target detection is made more difficult by a number of problems with underwater imagery, such as low contrast, color distortion, fuzzy texture features, and noise interference, which are caused by the limitations of the unique underwater imaging environment. In order to solve the above challenges, this paper proposes a multi-color space residual you only look once (MCR-YOLO) model for underwater target detection. First, the RGB image is converted into YCbCr space, and the brightness channel Y is used to extract the non-color features of color-biased images based on improved ResNet50. Then, the output features of three scales are combined between adjacent scales to exchange information. At the same time, the image features integrated with low-frequency information are obtained via the low-frequency feature extraction branch and the three-channel RGB image, and the features from the three scales of the two branches are fused at the corresponding scales. Finally, multi-scale fusion and target detection are accomplished utilizing the path aggregation network (PANet) framework. Experiments on relevant datasets demonstrate that the method can improve feature extraction of critical targets in underwater environments and achieve good detection accuracy.