Research On Target Recognition Lightweight Model Based On Deep Learning

Deep learning(DL),as an advanced technical means,is widely active in artificial intelligence,big data,voice recognition and other domains.Particularly in machine vision,the improvement of convolutional neural network drives the improvement of target detection.In the wake of the continuous progress of military technology,the form of war has developed from mechanisation and informatization to intelligent,and a large number of intelligent military equipment has emerged for enemy detection,strategic decision,intelligent operations and other aspects.The demand for miniaturization and intelligence of military equipment is also imminent.Target recognition model based on deep learning,therefore,not only need to consider the model parameters,the precision of model size,model,also need to comprehensively consider the resources of embedded devices,practical problems,such as volume,the parameters of the large scale and take up more space,target identification model to calculate the complex deployment to resource-constrained scene,deep learning oriented identification model shortcomings exposed.This dissertation explores compression and acceleration schemes for DL model,to obtain target identification model of high precision,high efficiency,promote academic work and practical application of close degree,high precision,low latency,lightweight model of low power consumption,further expansion of target recognition algorithm based on depth study of application scenarios.The research emphases and innovations of this paper involve the following three aspects:An IU-Cycle GAN data amplification algorithm is proposed to solve the problem of insufficient data in special domain.Firstly,the GAN network is replaced by Cycle GAN network in terms of data enhancement,which solves the problem that the result samples generated by traditional GAN network are not rich enough.Secondly,in order to improve the generation effect of generator model,U-Res Net structure is proposed combining the advantages of long and short connections of U-NET and Res Net,which effectively improves the generation effect of model.Moreover,there are many jump connections in U-Net,and channel mixing structure is introduced to further integrate deep information and shallow information.Improved U-Res Net generation.Finally,aiming at the problem that Cycle GAN training is not stable enough,the stability of wasserstein distance lifting model training is used to improve the generation effect of the model.The experiment fully proves that the IU-Cycle GAN data amplification algorithm can availably improve the diversity and authenticity of the data set and the accuracy of the model considering the trait of varied data sets.Aiming at the problem of YOLOv4 model’s large volume and redundant parameters,this paper proposes a target recognition model compression algorithm based on Ghost module.Firstly,in order to make the parameter distribution of the model more reasonable,the model was modified,and the feature pyramid pooling module is adopted at multiple stages of the feature pyramid,which can realize the fusion of local and global features at diverse scales,so as to broaden the receptive field and optimize the model.Secondly,Ghost convolution module is used to reconstruct YOLOv4 target recognition model.Plug-and-play Ghost module is very highefficiency,retains redundant features,can lifting the inference speed of the model and compress the volume of the model under the circumstance of guaranteeing the accuracy.Finally,the ultimate pruning compression algorithm of channel-convolution module is adopted,and the importance of different channels and convolution blocks is evaluated based on the batch normalization layer by sparse training.Based on this,the pruning operations of the model are accomplished,and the model is compressed in terms of network depth and width.Experiments bear out the meliority of the compression algorithm based on Ghost module,and real-time detection is realized on embedded platform.A knowledge distillation algorithm based on YOLOv4 target recognition algorithm was proposed to solve the problem that the model parameters were compressed greatly and the accuracy was damaged after extreme pruning.Firstly,channel-space attention is introduced as knowledge experience to transfer the attention information of teacher network to student network,so as to improve the identification performance of student network.Then,the multi-stage teaching assistant strategy was introduced to train the student network to reduce the model gap between the two knowledge distillation networks and improve the distillation effect.Finally,experiments are carried out to verify the effectiveness of the knowledge distillation algorithm based on YOLOv4 target recognition algorithm.