Research On Object Detection Algorithm Based On Feature Fusion And Dual Attention Mechanism

Object detection,as a popular research direction in artificial intelligence,is widely used in security surveillance,intelligent transportation and other fields.With the rapid development of deep learning,object detection technology based on deep learning has become an important research topic in this field,mainly using convolutional neural networks for accurate classification and localisation of targets in images.In the detection process,factors such as small targets in complex environments,variable light and different scales of objects often increase the difficulty of detection and reduce the detection accuracy.This paper focuses on the problem of detecting small targets in complex backgrounds and multi-scale target images to construct effective detection methods.The main work is as follows.(1)To address the problems of incomplete feature extraction and high miss rate of small targets due to complex background factors in the object detection process,a object detection algorithm with a dual attention mechanism for discriminant correlation analysis is proposed.The algorithm focuses on the optimization of the Faster R-CNN backbone network,and firstly,discriminant correlation analysis technique is introduced instead of the conventional feature fusion method to ensure the interaction between information by maximising the correlation between corresponding features in the feature set.Secondly,based on the advantages of the attention mechanism,a residual double attention mechanism is proposed to extract effective feature information.At the same time,a hybrid convolutional layer is designed to maximise the feature extraction performance of the network.(2)To address the problems of incomplete feature extraction and slow detection speed caused by multi-scale targets in the object detection process,a object detection algorithm with a channel-separated dual-attention mechanism is proposed.The algorithm mainly improves the Faster R-CNN+FPN backbone network by firstly adding a constructed detail extraction module to the original FPN network structure,which is used to alleviate the problem of weak feature information of small targets.Secondly,the channel-separated dual-attention network is designed to achieve effective extraction of deep-level features.Meanwhile,to further prevent the network from overfitting,group convolution and null convolution techniques are introduced to control the number of parameters in the network,and the scale change problem in target detection is effectively solved by optimising the loss function.The two detection algorithms proposed in this paper are experimentally compared with classical algorithms such as DPM,Fast R-CNN,Faster R-CNN,SSD,R-FCN and YOLOv1 on the PASCAL VOC2007,KITTI,Portrait and Pedestrian datasets.The experimental results show that the algorithms in this paper can efficiently recognize images with certain robustness and practical application value.In this thesis,there are a total of 43 figures,15 tables,and 92 references.