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

Object detection is an important research direction in the field of computer vision,which is mainly used for the recognition and localization of objects in images.At present,the object detection algorithm based on deep learning has become an important support for the implementation of artificial intelligence technology in the industrial field,and is widely used in security monitoring,industrial manufacturing,medical diagnosis and other fields,and has achieved good results.However,due to the complex and changeable scene,the large change of object scale,the dense overlap of objects and other factors,current object detection algorithms still have the problems of missing detection and low accuracy.In addition,most detection algorithms with high accuracy are difficult to deploy on mobile devices due to their large model size and high computational complexity.Therefore,the lightweight research of object detection algorithms is increasingly important.To solve the above problems,based on the existing YOLOv4 object detection framework,this thesis optimizes the detection performance by improving the feature extraction ability and feature fusion ability of the model,and carries out lightweight design of the detection model.The main research contents of this thesis are as follows:(1)Aiming at the problem of accuracy degradation caused by the changing scale and dense alignment of objects,an object detection algorithm based on adaptive weighted fusion module is proposed.The analysis is carried out from the perspectives of improving the ability of the backbone network to extract features and improving the ability of the neck network to utilize features.Firstly,an adaptive weighted fusion module is designed to change the summation of features in the residual blocks to adaptive weighted fusion,and fuses the proposed adaptive weighted fusion module into the residual block of the backbone network,so that the backbone network can focus on extracting features that are more useful for subsequent tasks.In addition,a cross-stage fusion path is added to the cross-scale feature fusion network,and the output features of the backbone network are connected with the features after feature fusion through skip connection,so as to make up for the information loss of features in the fusion process and enable the neck network to make full use of the feature information.The proposed algorithm is compared with the mainstream object detection algorithms on the public datasets,and the results show that the proposed algorithm has advantages in detection accuracy.(2)Aiming at the problems of large model size,high computational complexity and low accuracy after lightweight model,an object detection algorithm based on lightweight convolution and feature enhancement is proposed.In order to achieve a good balance between high accuracy and lightweight,the lightweight network MobileNetv2 is used as the backbone network and feature fusion is performed using depth-wise separable convolution in the neck of the network to further lighten the model.Then,the attention mechanism is introduced into the output position of the backbone network to enhance the effective information of the features.In addition,a shallow feature enhancement module for facilitating feature fusion is proposed and applied to the neck network to propagate information from deep features to shallow features in a shorter path to complement the semantic information of shallow features,so as to enhance the detection effect of small objects.Finally,the effectiveness of the model lightweight method and feature enhancement module is verified by experiments.The experimental results also show that the proposed algorithm has the characteristics of lightweight and good detection accuracy.