MSFC: Multi-scale Significant Feature Correction Method for Distorted Underwater Target Recognition

Abstract

The features of small-scale targets are less in underwater, and the complex underwater environment and light refraction cause the images to be distorted. These problems caused the AUVs difficulty to recognize distorted and multiscale targets in underwater. In view of the above problems, a multi-scale significant feature correction method was proposed for distorted underwater target recognition. Firstly, building a feature pyramid network through DenseNet backbone network to extract the multi-scale significant feature. Secondly, the method in this paper performs contrast correction of distorted image feature locations and recovers a clear grayscale image of the target. Finally, the method uses the extracted corrected multi-scale significant features for accurate recognition of multiscale underwater targets in images. Experimental results show that the method in this paper has better results in both image distortion correction and multi-scale targets recognition, with an accuracy improvement of 2.27% over existing methods.