Improve the semantic perception ability of CNN-based object detection systems with attention model

Convolutional neural networks(CNNs) based object detection methods are prone to be interfered with by background noise and cannot make full use of semantic information in the positive sample-choosing phase. To overcome these issues, first, we proposed an external module called “Semantic Information Attention Module (SIAM)”. The SIAM module can reduce the influence of features from non-target areas. Furthermore, we proposed a new positive sample selection strategy “Heat IOU”, and the “Heat IOU” standard will give candidate samples with more semantic information more weight in the positive sample selection phase. Note that SIAM and Heat IOU can be easily combined with existing convolutional backbones, which provide a flexible plug-and-play way to spur the object detection system to focus on the zones in the image with abundant semantic information in forward propagation and positive sample choosing phase. Experimental results demonstrate that our SIAM and Hot IOU successfully improve the performance of the object detection task. The code is available at: https://github.com/OZERO98/SIAM