| In the field of computer vision,the development of deep learning technology has achieved better performance than traditional methods.However,while continuously improving performance,deep learning models also bring shortcomings such as large network parameters,complex model structures,high computational costs,and large memory consumption.This makes running deep learning models on embedded mobile devices with limited resources Be difficult.Therefore,it is of great significance to study how to reduce the size of the deep learning model and reduce the amount of calculation,and deploy deep learning on the mobile terminal.In view of the above problems and characteristics,this paper studies the compression and optimization technology of convolutional neural networks,optimizes and accelerates the lightweight neural network model MobileNetV2,builds a screen scene recognition system under the Android system,and deploys it to enterprises based on actual projects.On mobile TV products,it provides more accurate,fast and stable TV scene recognition technology for the image quality adjustment function of embedded terminal devices.First,this article reduces the structure of the model by removing the 7-layer inverted linear residual block in the MobileNetV2 model,and the compression effect is to reduce the parameter amount of 0.9M and the calculation amount of 56 M.Then use the Slimming method to channel pruning the model,process each neuron as a channel,introduce a scale factor ? for each channel to act as a proxy for channel selection,apply an L1 penalty term to the scale factor,and use the Image Net data set to obtain For the pre-trained model,the pruning effect is to reduce the model size by 25%,and the accuracy is 69.9%.Finally,the pre-training model after channel pruning is changed to the last layer of softmax layer and the remaining weights are retained for transfer learning.The TV screen scene data set is used to fine-tune the training,and the accuracy rate is 70%.Then quantize the model through the training quantization technology provided by the MACE framework,compress the model size by 3-4 times,increase the running speed by 2-3 times,and reduce the accuracy rate by less than 0.1%.Finally,the trained model is transplanted to the Android mobile terminal through the MACE framework,and the Android application is developed.The running speed on the Android terminal is about 30.9ms,the recognition accuracy rate is about 71%,and the occupied memory is between 70~90MB.The rate is between,and the CPU occupancy rate is between;after the mobile phone test meets the project requirements,the model is combined with the picture adjustment function and deployed on the TCL 75P8 artificial intelligence network flat-screen TV terminal for TV picture enhancement.The model recognition time is 165 ms The recognition accuracy rate is about 75%,and the memory occupation is between 80~90MB.The final experimental results show that compared with the original model,the improved model can run steadily on embedded devices such as mobile phones and TVs.The efficiency of the model depends on the hardware and software configuration of the test equipment.This article reduces the model's lower accuracy rate while also greatly reducing the model parameters,size and calculation,and improves the model running speed,which is very important for embedded devices with scarce resources.The model has been successfully transplanted to the TV platform and docked with the picture adjustment function,and is currently advancing the productization of the project;the specific functional effects of the project have been demonstrated within the company and exhibited at the CES 2019 exhibition. |
PDF Full Text Request