Research On Object Detection Algorithm Based On Feature Fusion And Region Candidate

Object detection technology is an important research topic in computer vision and digital image processing.Its purpose is to identify the most attractive object from the image and reduce the consumption of human capital,which has important practical significance.However,since the object to be detected usually has the characteristics of background confusion,overlapping objects,and inconsistent scales,it becomes more difficult to detected in complex scenes.On the one hand,many researchers do not sufficiently consider the connection between shallow location information and deep semantic information in the feature pyramid,but only follow a simple chain aggregation structure to extract features,which can cause the problem of insufficient feature information expression,thus leading to the object detection task cannot accurately detect the category probability and location coordinates of the objects of interest;on the other hand,the region candidate frame is dependent on the ratio and size of the preset anchor frame,which causes the problem of unbalanced number of positive and negative samples,making it difficult to improve the robustness and feasibility of the multi-scale detection task in complex detection scenarios.To address the above problems,a feature fusion-based and region candidate object detection algorithm is proposed in this paper,and the main work is divided into two parts.In the first part,a novel feature fusion-based Feature Pyramid Network(Su FPN)is proposed.In the feature pyramid,both semantic information and location information are considered in a balanced manner.The deformable convolution is used in the lateral connection to expand the receptive field of the detector,and channel attention is used in the top-down path to effectively supplement and correct each feature information of one layer,while effectively increasing the correlation between adjacent layers.Thereby aggregating spatial and channel information to generate a discriminative feature pyramid with positional semantic interdependence,improving the accuracy of detection and classification.A large number of experiments on Su FPN on PASCAL VOC and COCO datasets show that Su FPN has a higher accuracy rate than methods such as FPN,PANet,CBAM,and Libra-RCNN.In the second part,an adaptive convolution-based region candidate network(ARRPN)is proposed.Considering that in real scenarios,positive and negative samples are often unbalanced.In order to get enough positive samples,a novel joint sampling approach is proposed to take centroid sampling and interaction ratio-based sampling in two refinement iteration stages to increase the number as well as quality of positive samples in the model,and an adaptive convolution module is used to increase the perceptual field of the model and initialize and learn the anchor center position and shape offset in the two stages,respectively.The deformable convolution is guided by the offset vector to ensure that the feature alignment principle is followed between the feature map and the anchor points.Extensive experiments with ARRPN on the COCO dataset show that ARRPN has higher recall,accuracy and generalization capability compared to methods such as Faster R-CNN,Iterative RPN+,GA-RPN and Cascade RPN.