Research On Two-Stage And Anchor-Free Object Detection Algorithm

As the cornerstone of image understanding and computer vision,object detection is the basis for solving tasks such as image segmentation,scene understanding,and object tracking.Although there are many kinds of object detection algorithms,according to whether the region proposal network is included,it can be divided into two-stage object detection algorithms and single-stage object detection algorithms.According to whether the anchor boxes are included,it is divided into anchor-based object detection algorithms and anchor-free object detection algorithms.The current object detection algorithm has achieved great improvement in accuracy and speed.However,due to the diversity of perspectives,multi-scale changes,the influence of complex scenes such as occlusion and illumination intensity,and the insufficient generalization ability of the model,object detection is still a challenging task.In response to the above problems,this paper mainly conducts in-depth research on building a better feature pyramid and avoiding the use of anchor boxes.The main research contents are as follows:(1)Since the essence of the anchor-based object detection algorithm is to use the sliding window for dense prediction,although the huge solution space can obtain a high recall,it is easy to obtain too many negative samples.Moreover,for images with object occlusion,it will also lead to the problem of semantic ambiguity caused by the high overlap of object centers.Therefore,this paper proposes an object detection algorithm that combines Bidirectional Feature Pyramid Network(Bi FPN)and Faster R-CNN.Bi FPN can not only alleviate the impact of the highly overlapping object centers,but also effectively solve multi-scale problems.(2)Since the anchor-based object detection algorithm will introduce hyperparameters related to the anchor box,and the predefined anchor box will make the model lack the generalization ability.In addition,when Bi FPN fuses different input feature maps,it does not consider their contribution to the output feature map.Therefore,this paper proposes an object detection algorithm that combines Weighted Bidirectional Feature Pyramid Network(WBi FPN)and anchor-free detector.The anchor-free object detection algorithm does not need to introduce hyperparameters related to predefined anchor boxes,so it does not need to spend extra time to find the optimal solution of these hyperparameters.And there is no need to set different anchor boxes for different datasets,so it has better generalization ability.WBi FPN introduces an attention mechanism on the basis of Bi FPN,which makes different input feature maps have different contributions to the output feature map,so that a better feature pyramid network can be constructed to further alleviate the impact of multi-scale problems.(3)Since the anchor-based detector selects the feature level for an instance based on the Intersection over Union(Io U)of the instance and all anchor boxes on each feature layer,and for an instance,it only utilizes one layer in the feature pyramid network.However,the anchorfree model does not have this constraint,so this paper proposes a weighting strategy of softselected feature pyramid level to assign the same instance to different feature levels for prediction.Thus,the feature pyramid network is more fully utilized.In addition,not all feature points in the feature layer corresponding to an instance contain useful semantic information,and the amount of information contained in each feature point is also different.Therefore,this paper proposes a weighting strategy of soft-weighted anchor points to re-weight the detection results of each feature point.Combining these two soft weighting strategies with WBi FPN further improves the detection accuracy of the model.