作者:\n\t张宁,于鸣,任洪娥,陶锐,赵龙\n
Authors:\n\tZHANG Ning,YU Ming,REN Honge,AO Rui,ZHAO Long\n
摘要:\n\t当前基于深度学习的目标检测技术得到了迅速发展,但小目标检测仍然是一个有待改善的难题。相比于大目标,小目标检测任务存在分辨率低、特征易丢失等特点,很多通用的目标检测算法不能直接迁移到小目标检测。特征金字塔融合能有效结合深层和浅层的特征,增强对小目标的检测性能,然而现有模型大都忽略了相邻层间融合时的信息不平衡问题。针对此问题,提出将有效融合因子的思想融入YOLO-v4的PANet结构,添加融合因子L-α控制深层向浅层传递的信息量,从而有效提高信息融合效率,增强YOLO-v4对小目标的检测能力。实验表明,加入了L-α的YOLO-v4模型,在TinyPerson数据集上平均精度APtiny50和APsmall50分别提高了2.14%和1.85%,在MS COCO数据集上平均精度AP和APS分别提高了1.4%和2.7%,且检测结果优于其他小目标检测算法,证明此改进方法对小目标检测有效。\n
Abstract:\n\tThe detection ability for small object is still need to be improved urgently in spite of the rapidly developing object detection technology based on deep learning at present.Compared with large objects, small object detection tasks hold drawbacks of low resolution and feature loss which leads to that many general algorithms cannot be directly applied to small object detection.The feature pyramid fusion can effectively combine the features of deep and shallow layers to enhance the performance.To solve the problem most models existingignoring the imbalance of information during the feature fusion between adjacent layers, it is proposed to integrate the idea of fusion factor into the PANet of YOLOv4, use the fusion factor L-αto control the amount of information transmitted from the deep layer to the shallow, so as to effectively improve the efficiency of information fusion and enhance the ability of YOLO-v4 for small objects detection.With the addition of L-αin YOLO-V4 model, the experiment results show that the APtiny50and APsmall50on the TinyPerson are improved by 2.14% and 1.85% respectively, while the APand APSon the MS COCO are separately increased by 1.4% and 2.7%.It is proved that this improved method is effective for small object detection with the evidence of better result than other small object detection algorithms.\n
PDF全文下载地址:
可免费Download/下载PDF全文
删除或更新信息,请邮件至freekaoyan#163.com(#换成@)