Ship detection of coast defense radar in real marine environment based on fast YOLO V4

Abstract

We propose an innovative method for ship detection in the real marine environment. Different from other high-resolution optical images and synthetic aperture radar images, ship detection of coast defense radar is quite challenging due to the complex background, sea state, and low resolution. To this end, we build a real dataset and propose an innovative detection method based on you only look once (YOLO) V4. Specifically, first, the lightweight architecture MobileNetV3 is introduced as the backbone feature extractor to accelerate the detection speed by compressing the parameters. Second, for better detection of small-size ships, the squeeze-and-excitation module is used to apply the attention mechanism to the channel. Meanwhile, the scaled exponential linear unit non-liner activation function replaces the rectified linear unit activation function of the MobileNetV3 shallow layer, which optimizes the convergence effect of the model. Third, an adaptive anchor-selection algorithm for the detection of ships with various shapes is designed. Compared with other well-established models based on convolutional neural network (CNN), including single shot multi-box detector, Faster region based-CNN, and you only look once version 4 baseline for detecting ships, our improved method yields impressive results in our dataset. After extensive testing, the mean average precision of the proposed method can reach 97.43%, with the detection time per frame reaching 38 ms.