Research On Compression And Application Of Target Recognition Model

With the development of society,people's material life and other aspects have made great progress.but people also pursue higher requirements in life and traditional industrial technology has been unable to meet people's high quality requirements for service.The existence of AI(artificial inteligence)can solve this problem well.The emergence of artificial intelligence will bring people into an intelligent society and its diverse functions,gorgeous services bring people a good experience.But because of its limitations,it is difficult to adapt to the higher requirements of intelligence.It is in this context,deep learning begin to attract people's attention because of its learning automation,better results and more favored by people.At present,deep learning has made great progress in target recognition and image segmentation.Intelligent devices with low price and low power consumption will be very popular in the future,but there is a big drawback of deep learning,that is,the requirement of hardware is too high.Deep learning algorithm requires a lot of computation,which makes it difficult to deploy to embedded devices with limited computing power and energy.Therefore,how to compress the deep learning model on the basis of keeping the original accuracy of the model as soon as possible has become a very challenging research topic.In order to solve this contradiction,this design aims to design a more intensive and less computational deep learning model,so that it can be deployed to embedded devices while ensuring the performance of deep learning and then deploy to home service robot.This paper focuses on how to compress the recognition model and how to apply it in practice.The main contents include:(1)Based on the current popular deep learning recognition model YOLO(You Only Look Once),this paper combines a variety of compression strategies,compresses the model on the basis of Google LeNet,and then transplants it to the Caffe framework for training,and finally obtains the final recognition model Compress-YOLO.Then,the model is deployed to the home service robot for practical test.The test results show that Compress-YOLO can accomplish the target recognition very well.(2)Based on the analysis of the model,it is concluded that the key of the compression model is to remove redundancy and make the model more compact.In order to solve this problem,this paper decomposes large convolution kernels into several convolution kernels of different sizes,increases the ability of extracting multi-scale features of the model,and adds 1×1 convolution kernels to the appropriate location,which reduces the dimensionality and increases the ability of cross-channel information integration of the image.At the same time,aiming at the phenomenon of over-fitting in training,this paper abandons the commonly used method of dropout layer and adopts the method of combination of Batch Normaliztion layer and Scale layer,which not only suppresses the over-fitting effectively,but also improves the training speed of the model.In addition,in view of the international common data set objectives and diverse environments,this paper establishes NEU-Family dataset,which enhances the recognition ability of the model to family goals by training on the dataset.(3)After completing the compression of the deep learning model,this paper carries out a specific application experiment for a specific home service robot project.The model is deployed on the NVIDIA Jetson TX1 development board to recognize the target in the home environment,which achieves good results and fully proves the validity of the compression model.