Research On Object Detection Method Based On Feature Fusion And Adaptive Attention

As an important research field of computer vision,object detection is currently widely used in fields such as self-driving,industrial manufacturing,intelligent medical care,and security.However,in object detection,there is still a problem that due to the complexity of the object detection scene and the large difference in the size of the object to be detected,the prediction branch usage features do not match the features required by the detection task,resulting in low detection network accuracy.Aiming at the above-mentioned problems in the object detection network,this paper proposes a object detection method based on multi-scale feature fusion and attention mechanism.The main work is as follows:(1)In the object detection task,there is a large size difference between the objects to be detected,which makes it difficult to detect some objects effectively.Based on the YOLO v3 network,using the characteristics of atrous convolution that can effectively improve the size of the receptive field of the network layer,constructing a multi-layer parallel atrous receptive field module;and fusing the bidirectional feature pyramid structure to construct the BR-YOLOv3 object detection network.The experimental results show that the improved RB-YOLO v3 has better test accuracy on the Pascal VOC test set than mainstream networks such as YOLO v3,SSD,Faster RCNN.(2)Aiming at the problem that the feature map extracted from the object detection network contains a large amount of redundant information,which is easy to affect the object positioning and classification tasks.This paper uses spatial attention and channel attention mechanism to construct a grouped hybrid adaptive attention module to improve the target detection performance of the network.The experimental results show that by using the improved attention mechanism,the detection accuracy has been improved on Retina Net,FCOS and other models.(3)In the flame detection task in the actual application scenario,there are many interfering targets,which are easy to cause false detection of the flame.By adding objects with similar characteristics to the flame as negative samples in the training set,a flame detection data set with practical application significance is constructed.The network obtained by training with this data set has good anti-interference performance and robustness in complex scenarios.At the same time,in the flame detection application scenario,the atrous receptive field module and grouped attention module are applied to the experiment,and the experimental results further show the effectiveness of the above-mentioned improved module for improving the detection accuracy of the object detection network.