Research On Two Stage Rotating Object Detection Algorithm Based On Anchor Free Frame

As a basic task of computer vision,object detection has received extensive attention,and rotating object detection is a branch of the field of object detection.For arbitrary and densely distributed extreme aspect ratio targets,the horizontal frame cannot accurately surround the target,resulting in a large amount of background redundancy in the label frame and serious overlapping of adjacent label frames,which will make it difficult to extract target features.However,using spinbox annotations can solve this problem.In the field of remote sensing target detection,the target is usually marked with a rotating frame and detected using a rotating target detection algorithm.Remote sensing target detection plays an important role in both military and civilian fields,and is widely used in resource surveys,urban planning,etc.Therefore,how to design a rotating target detection model with higher accuracy and faster speed is a research topic of great significance.This thesis will conduct in-depth research on this topic.In rotating target detection,the preset anchor frame scheme will cause a substantial increase in model calculations,and unreasonable anchor frame setting strategies will also lead to difficulty in model convergence and a decrease in detection accuracy;There are sensitivity,periodicity and boundary problems in the parameter representation method of rotating box based on angle regression.In order to solve the above problems,this thesis designs an anchor-free frame detection strategy based on key point sets,and each feature point predicts 9 key points for target calibration.In the training phase,the weak supervision method is used to constrain,and the four key points closest to the vertices of the annotation frame are selected as representative points to participate in the loss function calculation;in the inference phase,the minimum circumscribed rectangle algorithm is used to obtain the predicted rotation frame.Most of the two-stage models are designed with reference to Faster RCNN.In the RoI alignment stage,there are problems such as misalignment of features and difficulty in extracting rotation-invariant features.In the post-processing stage of the prediction frame,the violent screening of the NMS algorithm causes the loss of position information.In order to solve the above problems,this thesis designs a multi-feature point information fusion detection algorithm.After the region of interest is obtained in the first stage,multiple feature points in the region respectively predict the boundary points of the target and form key points with the representative points of the previous stage.Set to achieve fine adjustment of the target position.For the loss function,the focal loss is introduced to alleviate the long-tail effect,and the key points of the prediction that fall outside the annotation frame are penalized.Traditional convolutional neural networks do not have rotation equivariance,which will cause serious loss of target rotation features.In order to solve this problem,this thesis selects the Swin Transformer with more accurate rotation feature extraction as the backbone network through the visual analysis of the classification activation map of Swin Transformer and ResNet.At the same time,the deformable convolution DCN is introduced in the detection head to replace the traditional convolution,which effectively improves the detection accuracy of targets with severe scale changes and extreme aspect ratios.Finally,this thesis designs a two-stage rotating target detection model based on anchor-free frames,and confirms its excellent detection performance through a large number of simulation comparison experiments.Compared with the anchor frame-based single-stage model R3Det,the model designed in this thesis reduces the amount of parameters and calculations by 39.8%and 66.7%respectively,and increases the inference speed and detection accuracy mAP by 26.8%and 2.82 respectively;Compared with the two-stage models ReDet and Oriented RCNN based on anchor frames with excellent performance,the models designed in this thesis have reached the same level in detection accuracy,but the inference speed has increased by 45.6% and 25.2% respectively.