作者:赵一鸣， 王金聪， 任洪娥， 赵 龙
Authors:ZHAO Yi-ming, WANG Jin-cong, REN Hong-e, ZHAO Long摘要:摘 要:基于深度学习的小目标检测研究对于如小人脸识别、遥感图像检测等任务的优化与提升都具有极为重要的意义。但由于图像中的小目标所占像素较少，分辨率低，包含的特征信息不明显，现有方法对小目标的检测效果并不理想。针对此问题，提出一种基于反馈的特征融合网络ReFPN用于YOLOv4算法，两次利用骨干网络提取的原始特征层，加强小目标特征信息，对其进行更精确的位置回归。同时提出混合注意力机制Co-AM充分提取小目标的细节特征信息，抑制无效特征，进一步提高小目标的检测精度。实验结果表明，此文提出的方法使YOLOv4算法在MS COCO数据集上平均精度AP提高了1.9%，小目标平均精度AP_S提高了3.3%，检测效果优于现有小目标检测算法，证明了此文提出方法的有效性。
Abstract:Abstract:The research of small target detection is of great significance for the optimization and improvement of tasks such as small face recognition and remote sensing image detection. However, the small target in the image occupies fewer pixels, lower resolution, unobvious feature information, resulting in the effect that existing methods for small target detection is not ideal. To solve this problem, a feedback-based feature fusion network (ReFPN) for YOLOV4 algorithm was proposed.The original feature layer extracted from the backbone network is used twice to enhance the feature information of small targets and position regression performs more accurately. At the same time, the compound attention mechanism (Co-AM) was proposed to more fully extract the detail feature information of small targets, suppress invalid features, and further improve small targets’ detection accuracy.Experimental results show that the method improves the AP of YOLOV4 algorithm on MS COCO dataset by 1.9%, and the AP_S by 3.3%. The effectiveness of our method is proved, and the detection effect of small target detection is better than the existing algorithms.

PDF全文下载地址:

可免费Download/下载PDF全文