Object Detection Algorithm Based on Deformable Convolutional Networks for Underwater Images

The accuracy is low and the speed is slow of existing underwater images object detection algorithms, especially in variable scale forms on the same or different object (such as sea urchins, starfish, etc.). To address this problem, this paper propose an effective object detection algorithms based on deformable convolutional networks for underwater images. Firstly, we preprocess underwater images using convolutional neural networks and white balance to remove color deviation and increase the contrast, as well as improving the resolution. Then, in order to obtain a score map of the class information that has more representation, we rebuild ResNet-101 feature extraction sub-network using deformable convolution model to improve the feature extraction ability, and construct a deformable position-sensitive ROI pooling process with RPN. Finally, we apply softmax to regression and classification for feature information of each object in order to accurately output the classes and position. On testing dataset, the subjective and objective experimental results show better detection effect, and the mean average precision (mAP) is 90.3%, which is higher than the comparison algorithm. The experimental results show that the proposed algorithm has better detection accuracy and effect for object detection of underwater images.