A Ship Detection Method in Complex Background Via Mixed Attention Model

With the development of deep learning, recent object detection methods have made considerable progress. Unlike other simple visual object detection problems, it is more difficult to detect ships in a complex background. Nearshore vessels are always confused with background objects. When the ship target is small, the feature information in the background will greatly interfere with the information extraction and learning of the target area of the ship by the convolution neural network, resulting in the problem of ship sample imbalance. In response to this problem, this paper proposes a mixed attention model to decrease the difficulty of ship detection, which is composed of pixel attention model (PAM) and feature attention model (FAM). The PAM structure is a generative adversarial network designed to prove the sensitivity of the target area without extra manual works. FAM is a convolution network designed to increase the utilization rate of useful features. While MAM is not a fixed structure, it could be implanted into almost any object detection and classification networks. Meanwhile, PAM is for the original image preprocessing part and FAM is always added to the position of low-level features to high-level features. Following experimental results show that using MAM method into Yolov3, the mean average precision increased by 2.2% when achieved 0.975, which effectively improve the accuracy of ship detection in complex background.

Automated summary

This paper proposes a mixed attention model consisting of pixel attention model (PAM) and feature attention model (FAM) to improve the accuracy of ship detection in complex backgrounds. Experimental results show that using this method with Yolov3 increases mean average precision by 2.2%, achieving 0.975.