Lightweight Feature Extraction Network In Edge Computing Environment

In the field of computer vision,feature extraction is a crucial task that involves extracting both shallow and deep semantic features from images.This process enables subsequent algorithms to operate more efficiently,making it one of the most significant challenges in the field.In recent years,deep learning has become increasingly prevalent in various fields.However,the large number of parameters in deep neural networks can be computationally intensive,making it difficult to deploy on edge devices.As a result,there has been a rise in the development of lightweight feature extraction networks specifically designed for edge computing environments.However,these studies have not fully taken into account the unique characteristics of images in edge computing scenarios.Additionally,even after model optimization,the requirements for device resources remain high,leading to suboptimal network practical applications.This thesis proposes a feature extraction network that has been optimized using techniques such as parameter pruning and network structure transformation.The research focuses on the following specific contents:(1)Aiming at the problems that the existing networks do not adapt to the image characteristics of edge scenes,do not effectively process the detail information in images and have insufficient feature fusion,this thesis proposes a feature extraction network model SFPN that can better utilize the different depths and detail semantic information of the network.Firstly,the S-soft unit is designed to process the detail information of the feature map.Secondly,this thesis proposes a new secondary fusion strategy that combines the features of the middle layer are fused with the lower and higher layers respectively,and with this,SFPN Block is proposed as an auxiliary classifier,which participates in the final prediction by weighting,fully considering the prediction information of different layers of the network.Finally,a joint loss function is introduced for training,so that the feature extraction ability of the network for targets with large scale changes and small sizes is improved,and it is more adapted to the edge computing scenario of detection requirements and improve the accuracy of feature extraction.(2)A lightweight version of SFPN,Tiny-SFPN(TSFPN),is proposed to address the problems of large number of parameters,large computation and slow inference speed of SFPN model,and low real-time operation on edge devices.Firstly,batch normalization(BN)is performed on each convolution of the model during the training process to standardize the network parameters,and the parameter γ is obtained through training to facilitate subsequent judgment of channel importance.Secondly,a Judge discriminator is designed and multiple criteria are used to determine the cropping priority.Finally,the SFPN model is cropped and retrained in multiple rounds using the channel pruning method,which allows the model to reduce the parameters and computation while maintaining the feature extraction accuracy,enhance the operation speed,and meet the real-time requirements of image processing tasks in edge scenes.(3)To address the problem of limited inference speed and model performance of SFPN model in the deployment phase,a single-branch lightweight feature extraction model DRep-SFPN in the network deployment phase is proposed.On the one hand,the multi-branch of SFPN model is replaced with a single-branch structure that can be reparameterized to improve the inference speed of the model and reduce the number of parameters and computation.On the other hand,the conventional convolution operation is replaced by dynamic convolution to compensate for the degradation of model performance caused by the structural transformation while keeping the model lightweight,to find a balance between the speed and accuracy of feature extraction,and to further optimize the overall performance of the feature extraction model in the deployment phase.This thesis conducts experiments on three classical computer vision tasks: image classification,target detection,and semantic segmentation.The experiments use the Pascal VOC general dataset,Visdrone2018-DET,AI City Challenge target detection,and Cityscapes semantic segmentation dedicated datasets to verify the degree of performance improvement of the proposed feature extraction network.The experimental results demonstrate that the SFPN model,along with the channel pruning technique and network structure replacement method proposed in this thesis,can effectively adapt to the demands of edge scenes.These methods improve the accuracy and speed of feature extraction,and are shown to be progressive and effective when compared to similar models and techniques.