基于注意力机制的NMS在目标检测中的研究

Research on non-maximum suppression based on attention mechanism in object detection

非最大值抑制算法是目标检测中选择定位准确框的主要算法。针对算法仅以分类得分作为标准可能去除得分低但定位准确的预测框,并且对于有遮挡的情况更不友好的情况,提出了A-NMS方法。该方法将注意力机制融入非极大值抑制算法中,利用位置信息与框的得分信息相结合来调整框的最后得分。此外,还提出了改进的基于距离的交并比损失函数,重新定义了损失项,并将其引入到非极大值抑制中代替IOU来计算框间的交并比。最后将两种改进算法融合到3种经典的目标检测中,在Pascal-VOC 2012和MS-COCO 2017数据集上对上述两种算法进行了验证,结果表明检测精度得到了1%~2%的提升。

Non maximum suppression algorithm(NMS)is the main algorithm to select the accurate positioning box in object detection.The algorithm only takes the classification score as the standard,which may remove the prediction frame with low score but accurate positioning,and is more unfriendly to the situation with occlusion,A-NMS method is proposed,which integrates the attention mechanism into the non maximum suppression algorithm,and adjusts the final score of the box by combining the position information with the score information of the box.In addition,an improved distance based intersection union ratio loss function is proposed,the loss term is redefined,and it is introduced into non maximum suppression to calculate the intersection union ratio between frames instead of IOU.Finally,the two improved algorithms are integrated into three classical target detection.The above two algorithms are verified on Pascal-VOC 2012 and MS-COCO 2017 data sets.The results show that the detection accuracy has been improved by 1%~2%.