Research On Neural Network Compression For SSD-ResNet

Unmanned technology requires a camera to capture the surrounding environment,and machine vision is using visual products to sense the surrounding environment.Machine vision converts the ingested target into a picture signal that is sent to the processor for image processing to obtain specific information about the target.In machine vision,object detection is a basic algorithm whose goal is to obtain the position and classification information of an object from image information.During the actual object detection process,vehicle control in unmanned driving must be responsive and precise,and the moving vehicle is a mobile platform,which respectively provide fast and accurate image processing and low power consumption for processing.Up to now,the traditional machine learning algorithm for target detection has been gradually replaced by neural network algorithm.SSD-ResNet object detection network is one of the state-of-the-art detection algorithms,which has good performance in speed and accuracy.This paper will analyze the network and consider how it can be improved to fit low-power computing devices.When transplanting a neural network in a low-power computing machine,it is limited by storage and calculate ability,so the neural network must be compressed while taking into account the network performance.In order to solve the reqirement of compressing SSD-ResNet network and taking into account its performance,this paper conducted the following research: Firstly,several object detection neural networks were compared and analyzed to obtain the main idea of SSD algorithm and the main network replacement method.Combined with the structure of the ResNet network,the SSD-ResNet network was structured.Through the statistics,it was determined that the main work of compressing SSD-ResNet is the compression of the convolutional layer.The mAP was used as an evaluation standard for SSD-ResNet performance.Secondly,the neural network compression algorithm was researched.The three types of neural network compression methods were compared to determine the quantization method as compressing of the SSD-ResNet network.After comparing the quantitative strategies,it was determined that the values in the network convolutional layer are quantified using dynamic fixed points,and the values in the strategy were analyzed in detail.Then,a single-layer quantization process was designed according to the quantization strategy.According to the single-layer quantization process and the network calculation step,the network layer importance calculation method and the neural network quantization method with the mixed bit width based on importance are proposed.A retraining method for layerby-layer quantization and adjustment of neural networks is proposed as well.Finally,the proposed method was combined with the Caffe framework.The SSD-ResNet network with mixed bit width quantization is obtained through the quantitative experiment and retraining experiment of the whole network.The compression rate of the quantized network weight is 18%,while the accuracy is reduced by only 1.7%.