Research On Target Detection Algorithm Based On Bidirectional Feature Fusion

Object detection is one of the three core tasks in computer vision,which has important practical application value and research significance.In recent years,the development of deep learning has promoted the progress of target detection.At the same time,the progress of target detection has made the rapid development of the whole computer vision field,and promoted the vigorous development of autonomous driving,remote sensing monitoring and other industries.Therefore,target detection technology has great academic research value.Based on this,this paper mainly focuses on how to improve its detection performance.In target detection,feature fusion is used in many literatures.In the feature fusion process,high-level features have richer semantic information than low-level features,so the fusion method can extract features rich in semantic information and detail information,and further extract features that are more conducive to target detection.However,when using feature pyramid to detect objects of different scales,many literatures ignore the relationship between high-level information and low-level information,resulting in poor detection effect.In addition,for some small targets,it is easy to miss detection.In view of the above mentioned problems,the paper studies and analyzes the target detection on the general data set and remote sensing data set respectively,and puts forward an algorithm to effectively improve the detection accuracy of the general data set and remote sensing data set.The main achievements of this paper are as follows:(1)A target detection model combining bidirectional feature fusion and attention mechanism is proposedIn view of the problem that traditional algorithms tend to ignore the characteristics of small targets when detecting targets,this paper firstly analyzes SSD algorithm and finds that SSD has the problems of low efficiency and slow detection of small targets.Therefore,the SSD model has been improved.First of all,the characteristics of deep SSD model using bilinear interpolation method to enlarge the additive blend with the characteristics of shallow layer,and then get the top-down characteristics after the fusion layer,then based on the bottom-up features fusion,namely the characteristics of top-down merged down sampling method and the characteristics of deep figure for fusion.The semantic information of the fused feature image is enhanced by weighted updating of the fused feature image channel with attention mechanism.Finally,this paper improves the division strategy of positive and negative samples.When dividing positive and negative samples,the attributes of the target are fully considered,and the anchor related to the detected target is retained as the positive sample.Such improvement can reduce the target missed detection rate to a certain extent.Finally,the validation on VOC2007 and VOC2012 data sets shows that the algorithm proposed in this paper has higher average accuracy than the current mainstream algorithm.(2)Proposed the remote sensing target detection algorithm combining feature fusion and improved AnchorIn the process of remote sensing image target detection,it is difficult to extract the information related to the target,there are many missed targets,and the detection accuracy is low.To solve these problems,this paper proposes a remote sensing target detection algorithm combining feature fusion and improved anchor frame.First of all,the combination of image enhancement methods to increase the number of small target,at the same time,for the loss of detail information in the process of extracting small target features from the backbone network,the feature capture network is added to extract rich detail information.secondly,using per-pixel addition,the characteristics of top to lower and lower to the top of fusion,and extracted with characteristic fetching network details for further features fusion,semantic information and detail information to enhance detection layer,more small target feature extracting;Finally,in order to increase the number of small target positive samples,the proportion of Anchor is adjusted,and the matching strategy of positive and negative samples is improved.The algorithm proposed in this paper was verified by experiments on aerial remote sensing data sets,and the mean average precision(m AP)was 87.9%,which was 9.7% higher than the original SSD algorithm on UCAS-AOD data sets.The results further demonstrate the effectiveness of the proposed algorithm.