CS-DCFPN: An Object Detection Model

In practical application scenarios, the objects to be detected are characterized by a large number, irregular shape, non-uniform size and dense distribution, etc. Traditional object detection models are prone to the problem of object leakage or multiple objects in one anchor when dealing with the scenarios mentioned above. Based on the above problems, a new model for object detection named as Convolutional Self-attention and Deformable Convolutional Feature Pyramid Networks (CSDCFPN) is proposed. In this model, on the one hand, the SIAGIM module, which can capture both local and global semantic information, is proposed to effectively improve the detection accuracy of the model; on the other hand, the Deformable Convolutional Feature Pyramid Networks (DCFPN) is proposed for the problems of dense distribution, irregular distribution, and different sizes of the objects in the actual scene, in order to further improve the model to extract the irregularly shaped and irregularly distributed object features, thus improving the detection accuracy of the model. Finally, the CSDCFPN is validated on the wood detection dataset constructed based on realistic scenarios and the public dataset COCO2017. The experimental results show that the CS-DCFPN object detection model works optimally regarding mAP and recall rate.