Complex sea surface ship target recognition based on lightweight YOLOv4 and KCF

DOI:	10.11809/bqzbgcxb2024.11.001
Keywords:	lightweight YOLOv4 network; KCF algorithm; complex sea surface; ship target recognition; bidirectional equalization; feature map
Abstract:	For long distance or small size ship targets, the lightweight YOLOv4 model can provide efficient and real time feature extraction and reduce the demand for computing and storage resources, which is very important for long term maritime missions or mobile devices.And single scale feature extraction is easy to lead to large errors in recognition results.Therefore, a method of ship target recognition on complex sea surface based on lightweight YOLOv4 and KCF is proposed.First of all, the image of small ships on the sea surface is balanced in both directions to highlight the details of the image.Secondly, a lighter YOLOv4 network is designed to extract the feature map of ship targets from three different scales, so as to capture the position and dynamic changes of ship targets more quickly.Finally, the KCF algorithm combined with similarity threshold is used to screen out the target pixels, construct the ship target image, and complete the ship target recognition.The experimental results show that the image quality of the research method is improved after bidirectional equalization, with the highest SNR of 35.4 dB and the maximum SSIM of 094.The effect of lightweight feature extraction is ideal, and the minimum time complexity of feature extraction is 1.2 s; Compared with YOLOX S algorithm and cascade network method, the proposed method can identify all the ship targets with an accuracy of 100%.The maximum frame rate of the proposed method is 49.6 frames/s, which is 84.40% and 192.31% higher than that of YOLOX S algorithm and cascade network method, respectively.Therefore, it shows that this method can identify ship targets on complex sea surface more accurately.
Issue:	Vol. 45 No. 11 (2024)
Published:	2024-11-30
PDF HTML