Key Laboratory of Advanced Process Control for Light Industry Ministry of Education, Jiangnan University
The National Natural Science Foundation of China (General Program, Key Program, Major Research Plan)
现有的目标检测框架中，浅层弱分类能力是制约着网络精度进一步提高的关键. 针对上述问题，本文提出了基于浅层定位信息的动态细化检测网络.该网络在单阶段算法的基础上，通过增加多连接模块以增强浅层特征，同时去除浅层的分类操作以最大保留浅层的定位结果，并将其作为候选框送入深层网络. 深层网络通过使用引入自适应因子的感受野模块构建特征金字塔，以获得丰富的语义信息用于对浅层的回归结果进行判别和微调.最后设计基于自注意的可变形卷积头，通过对候选框的偏移以自发进行定位校准，使得网络获得精确的检测结果.在PASCAL VOC和MS COCO数据集上的实验结果，表明本文网络结构实现了优异的检测精度.
In the existing detection framework, the weak classification ability of the shallow layer is the key that restricts the further improvement of network accuracy. In order to solve the problem, a dynamic refinement detection network based on shallow positioning information is proposed. Based on single-stage algorithms, the network enhances the features of the shallow layer by adding multiple connection modules and removes the classification operations of the shallow layer to retain the location results of the shallow layer to the maximum. The location is used as the default boxes of the deep-level network. The deep level network is constructed by using a receptive field module with adaptive factors to obtain rich semantic information for the discrimination and fine-tuning of the regression results from the shallow layer. Finally, the designed deformable convolution head based on self-attention can automatically calibrate the position by shifting the detection box, which helps the network obtain accurate detection results. The experimental results on PASCAL VOC and MS COCO datasets show that the proposed network architecture achieves excellent detection accuracy.