Research On Target Detection Based On Multi-scale Feature Fusion

At present,with the continuous development of the theory and technology in the field of computer vision,target detection is an important part of its subject,and it has a very important position in many fields such as video surveillance,image retrieval,and smart home.How to improve the detection speed and detection accuracy of the target detection algorithm has become a research hotspot in the computer vision neighborhood.Traditional target detection methods usually use hand-designed features,and their detection speed and algorithm robustness are not high.The target detection algorithm based on deep learning directly generates features by using convolutional neural networks.Although the disadvantages of manual design features are avoided,the accuracy of detection is still to be improved.Therefore,how to design an efficient target detection algorithm with high detection accuracy is worthy of further study.The thesis studies the traditional target detection algorithm and the deep learning-based target detection algorithm,and proposes a new target detection algorithm based on deep learning.The target detection algorithm of this paper has two bright spots compared to other target detection algorithms based on deep learning.The first one is to improve the NMS algorithm.When the improved NMS algorithm uses the highest confidence detection frame to suppress other detection frames,instead of directly deleting other detection frames,the appropriate function is used to reduce its confidence.It can continue to participate in the subsequent suppression process.The improved NMS algorithm improves the quality of the suggestion box without adding any additional hyperparameters and algorithm complexity.The second bright spot is to propose a feature extraction network based on multi-scale feature fusion,and then construct a new target detection algorithm.In order to solve the contradiction between resolution and semantics in the feature extraction process,the algorithm proposes a multi-scale feature extraction network that performs upsampling of the feature map generated by this layer and merges with the upper feature map.The network can generate a feature map with high resolution and high semantic information,and the target detection algorithm constructed by using the feature map can improve the accuracy of target detection.In order to verify the detection effect of the proposed new target detection algorithm in the actual system,the algorithm and other current mainstream algorithms are used to conductcomparative experiments on the public data set.The experimental results show that the proposed target detection algorithm based on multi-scale feature fusion has better algorithm performance than the mainstream target detection algorithm Faster R-CNN,and the target detection accuracy is improved.