Research On Object Detection Algorithm Based On Anchor-free

The object detection technology is mainly used in the fields of automatic driving and intelligent monitoring.It requires the computer to locate the target of interest in the image containing multiple targets and identify its category,which has both academic research and engineering application value.The object detection methods based on deep learning are mainly divided into Anchor-based detectors and Anchor-free detectors.Due to the shortcomings of Anchor-based detectors such as poor versatility,positive and negative samples are easy to be unbalanced,this thesis mainly focuses on Anchor-free object detection,it does not depend on the pre-defined anchor boxes and avoids the complicated calculations related to the anchor boxes.The specific research contents are as follows:(1)Aiming at the problem that the adaptability of complex scenes is weak and the targets are easily interfered by background similarities,we designed an object detection model based on global context attention and feature fusion named AMFDet.The AMFDet model is a single-stage object detection algorithm based on Anchor-free.The model uses a pixel-by-pixel prediction method,does not rely on pre-defined anchor boxes or region proposal,and learns the context information in the image through a global context attention mechanism to make it pay attention to the target area,reduce the interference of the background of the feature map and the negative sample information.The AMFDet use the balanced feature pyramid to perform feature fusion to extract more feature information,and introduce focal loss and GIo U loss for classification and regression.The experiments on the PASCAL VOC and MS COCO data sets prove that the method in this thesis can effectively improve the performance of the model under the premise of ensuring the detection speed.Compared with the basic model FCOS,the m AP value of AMFDet model on the PASCAL VOC 2007 dataset has increased by 0.8%,and the AP value on the COCO dataset with complex scenes has increased by 1.2%.(2)Aiming at the problems of low detection accuracy and easy to miss detection of small targets in the detected images,we designed a high-resolution object detection model based on a guided anchoring region proposal network named HRGAD.HRGAD is a two-stage object detection algorithm based on Anchor-free.The model does not need to define the pre-defined anchor boxes but guides the generation of the anchor boxes through image features and predicts its position and shape.Extracting features through a high-resolution neural network can learn rich high-resolution representations,improve the spatial accuracy of small object detection,use balanced feature pyramid to improve the problem of insufficient representation capabilities of shallow small targets in the network.It use guided anchoring region proposal network to generate high-quality proposals,adopt bilinear interpolation Ro I Align technology avoids the problem of precision mismatch,and finally performs classification and regression to obtain the final detection result.The HRGAD model was tested on the PASCAL VOC and MS COCO datasets.The m AP value on the PASCAL VOC 2007 test set reached 83.9%,the AP value on the COCO 2017 test set reached 43.6%,the average precision of the small target APS value reached 25.8%,and the average recall rate of small targets ARS value reached 41.7%,indicating that the accuracy and recall rate of this model on small object detection tasks were significantly better than other models.