Research On One Stage Object Detection Algorithm Based On Anchor-free

In the field of computer vision,object detection is a basic problem,which has applications in areas such as autonomous driving,traffic monitoring,Unmanned Aerial Vehicle scene analysis,and robot vision.In recent years,with the rise of deep learning technology,object detection technology has achieved rapid development.Most of the deep learning object detection method are based on anchor boxes.In this type of method,the detection result is sensitive to the hyperparameters involved in the anchor,and it is easy to cause imbalance between positive and negative samples.in recent years,Anchor-free object detection methods have begun to emerge.This type of method abandoned the design of anchor,thus avoid the disadvantages mentioned above.In order to improve the accuracy of Anchor-free object detection method,this article research on these methods:1)In the convolutional neural network,the operation of convolution is fixed,and there is no mechanism to adapt to the geometric deformation of the object.In response to this problem,this thesis proposes to use deformable convolution to replace part of the convolutional layer in the backbone network,so that the sampling points of the convolutional layer have additional offsets,which can adapt to the deformation of the object.Next,this article studies the centerness used to surpress low quality positive samples.Aiming at the problem that centerness cannot intuitively reflect the degree of sample point deviation from the center point,this thesis proposes a new definition of centerness and named it distance-centerness.Distance-centerness uses the distance between sample point and center point to measures the degree of sample point deviation from the center point.Compared with the centerness,it can more intuitively reflect the degree of sample point deviation from the center point.Finally,this thesis uses the SoftNMS algorithm to reduce the degree of suppression of positive samples,and reduce the situation that positive samples are eliminated when occlusion occurs.2)In the field of object detection,in order to detect objects of different sizes,feature pyramid are usually used to generate multiple feature maps,and multiple feature maps are used to detect objects of different sizes respectively.In order to fuse features from different hierarchical feature maps,the feature pyramid adds an upsampling process so that low-level feature maps can obtain high-level features.This thesis proposes a new feature pyramid,which fully fuses features through top-down and bottom-up sampling processes to generate higher-quality feature maps.This thesis compares the method proposed in this thesis on the PASCAL VOC dataset and MS COCO dataset.The method proposed in this thesis has a significant performance improvement when compare with mainstream object detection methods.This article also conducted ablation study to ensure the effectiveness of the proposed method.