【摘要】 文中提出一种融合多头注意力机制、ROIAlign和Soft-NMS的FasterR-CNN目标检测算法，旨在解决原始Faster R-CNN目标检测网络中存在的检测精度低、漏检、误检的问题。首先，为了提高Faster R-CNN的感知能力，提取特征图中的重要特征并降低对无关特征的提取，在网络中嵌入注意力机制；接着，针对共享全连接层的降维操作导致的一些区域的细节信息被忽略，造成局部信息的丢失，采用一维卷积代替共享全连接层实现权重计算的任务，以捕捉更广泛的空间信息；然后为了提供更丰富的特征表达能力，在注意力机制中引入多头机制分别对特征的不同部分进行重要性的加权；为了减少在特征提取时原图信息的丢失，使用ROI Align替换ROI Pooling算法；最后，在算法后处理中引入Soft-NMS替换传统非极大抑制（NMS）算法以减少漏检和误检情况。实验证明，改进后的Faster R-CNN目标检测网络对感兴趣目标的定位能力得到提高，漏检和误检情况减少，平均检测精度得到显著提升。更多还原

【Abstract】 A faster R-CNN object detection algorithm that integrates multi-head attention mechanism, ROI(region of interest)align and soft-NMS(non maximum suppression) is proposed. This algorithm aims to improve the detection accuracy and eliminate the missed detections and false detections in the original faster R-CNN object detection network. In order to improve the perception ability of faster R-CNN to extract important features from feature maps and reduce the extraction of irrelevant features, an attention mechanism is embedded in the network. In response to the dimensionality reduction operation of the shared fully-connected layer, which leads to the neglect of detailed information in some areas and the loss of local information, one-dimensional convolution is used instead of the shared fully-connected layer to achieve the task of weight calculation, so as to capture broader spatial information. In order to provide richer feature expression capabilities, a multi-head mechanism is introduced into the attention mechanism to weight the importance of different parts of the features. In order to reduce the loss of information in the original image during feature extraction, ROI align is used to replace the ROI pooling algorithm. Finally, soft-NMS is introduced in the algorithm post-processing to replace the traditional NMS algorithm to reduce the missed and false detections. The experimental results show that the improved faster R-CNN object detection network improves the localization ability of interested objects effectively, reduces the missed and false detections, and improves the average detection accuracy significantly.更多还原