融合图像显著性的YOLOv3船舶目标检测算法研究

针对复杂水面环境下的船舶目标检测问题，运用融合图像显著性的YOLOv3船舶目标检测改进算法以提高检测能力。该算法基于Darknet-53网络模型，根据水上船舶特点，融合非极大值抑制算法Soft-NMS和显著性检测算法FT思想，进一步优化最终检测以达到更准确的效果。用Soft-NMS算法替换原有NMS算法，使得算法对小目标和重叠目标检测效果明显提升。融入FT算法对船舶图像局部细节作进一步细化，使得包围盒回归更加准确。在建立的数据集上进行训练与测试，实验结果表明，改进方法比原始方法准确率提高4%，达97%，检测速度提高10帧/s，达30帧/s，表明改进算法有效提高了船舶目标检测精度，且加快了检测速度。

Research on YOLOv3 Ship Target Detection Algorithm Based on Image Saliency

Aiming at the problem of ship target detection in a complex water surface environment， we employ an improved YOLOv3 ship target detection algorithm that incorporates image saliency based on the Darknet-53 network model. This method is based on the characteristics of watercraft， combining the non-maximum suppression algorithm Soft-NMS （Soft Non-Maximum Suppression） and the significance detection algorithm FT （frequency-tuned salient region） to further optimize the final detection to achieve a more accurate effect. The Soft-NMS algorithm is used to replace the original NMS algorithm， so that the algorithm can significantly improve the detection effect of small targets and overlapping targets. The FT algorithm is incorporated to further refine the local details of the ship image， making the regression of the bounding box more accurate. Training and testing on the data set established in this article， the experimental results show that the accuracy of the improved method is 4% higher than the original method， reaching 97%， and the detection speed is increased by 10 frames/s to 30 frames/s. It shows that the improved algorithm effectively improves the accuracy of ship target detection and speeds up the detection speed.