Improved YOLOv3 Algorithm for Ship Target Detection
Liankai Chen, Bangyu Li, Qi Liang · 2020
To solve the problem of ship target detection in complex surface environment, a real-time Darknet-53 network model based on a deep learning framework is incorporated with the improved YOLOv3 algorithm. This method combines a Soft Non-Maximum Suppression (Soft-NMS) algorithm and a Frequency-Tuned (FT) salient region detection algorithm to detect the features of ships on water. By replacing the original NMS algorithm with Soft-NMS algorithm, the detection effect of the algorithm for small target and overlapping target is improved obviously; Incorporates the saliency region features obtained by the FT algorithm based on frequency adjustment. The saliency region features obtained by FT contain more overall information of the target. Experimental results demonstrate that compared with the original YOLOv3, the detection accuracy and speed of the proposed method are considerably improved.