基于自深向浅特征融合的小目标检测方法

Small Object Detection Method Based on Deep to Shallow Feature Fusion

针对SSD算法各特征层关系未充分利用导致浅层特征层缺乏语义信息的问题,为提高对小目标的检测能力,提出一种自深向浅特征融合的小目标检测方法DTS-SSD(Deep to Shallow SSD)。使用BiFPN特征融合模块对特征多次提取获得多尺度语义信息;利用深层特征融合模块减少深层特征层缺失的小目标空间位置信息;构建1条自深向浅的特征融合路径来增强浅层特征层的语义信息;应用注意力机制学习特征图通道间的重要性。通过在PASCAL VOC2007测试集进行实验验证,mAP(Mean Average Precision)值达到80.1%,对目标的mAP较原SSD算法提高2.9%,该算法可行有效。

Aiming at the problem that the relationship between the feature layers of the SSD algorithm is not fully utilized,a small object detection method DTS-SSD(Deep to Shallow SSD)from deep to shallow feature fusion is proposed to improve the detection ability of small targets.The BiFPN feature fusion module is used to extract multiple features to obtain multi-scale semantic information;The deep feature fusion module is used to reduce the missing small object spatial position information of the deep feature layer;A feature fusion path deep-to-shallow is constructed to enhance the semantic information of the shallow feature layer;Attention mechanism is applied to learn the importance of feature graph between channels.The results in PASCAL VOC2007 test set show that the mAP(Mean Average Precision)value reaches 80.1%,and the mAP of the object is increased by 2.9%compared with the original SSD method,it is proved that the algorithm is feasible and effective.