Research And Application Of Embedded Residual Network Acceleration Method In Tumor Detection

Convolutional neural network has made many outstanding achievements in the field of tumor detection,among which the deep and excellent residual network can effectively guarantee the accuracy of tumor detection.However,the size of conventional residual network parameters is very large,and there are shortcomings of slow reasoning and difficult deployment on low power and low cost embedded devices.Therefore,it is of great theoretical and practical significance to study how to use residual network acceleration methods to realize accurate detection of tumor cells with low computation,low consumption and low cost on embedded devices.The network acceleration methods can effectively reduce the number of residual network parameters,but the traditional acceleration methods are limited to single network structure optimization and training model compression,so the acceleration effect is restricted.In this paper,a general embedded residual network acceleration method is designed which combines network structure optimization and model compression.The main research contents are as follows:Firstly,a residual structure with fewer initial parameters is designed and an accelerated embedded residual network is constructed.Preliminary experiments found that the original residual network can guarantee the accuracy of tumor detection,but the two residual structures that build them have more initial parameters.To solve this problem,a residual structure with optimization acceleration is designed in this paper,which is divided into two layers.In the first layer,the conventional convolution is decomposed efficiently by means of depth separable convolution and linear optimization The second layer uses the conventional convolution for feature extraction.The embedded residual network built with this structure has fewer initial parameters than the original residual network.The experimental results on Cifar 10 data sets show that the embedded residual network can reduce the initial parameters by more than 40% while maintaining the detection accuracy.Secondly,a hybrid compression acceleration method is designed,which can effectively reduce the complexity of embedded residual network model.Embedded network model which is generated by training still exists certain network redundancy,in order to solve this problem,this paper designs a kind of hybrid compression speed method,the method accelerates the network model from three independent direction including the parameters,computing and storage: The first is to cut out the convolution kernel and feature graph which have the weaker feature extraction ability.The second is merging the calculation of convolution layer and BN layer.The third is quantifying high precision floating point data.The experimental results on Cifar10 data sets show that the hybrid compression acceleration method reduces the storage of embedded residual network model by 86.3%,accelerates the speed by 69.1%,and reduces the precision loss.Finally,combined with the above two optimization methods,an efficient and universal embedded residual network acceleration detection method is designed and applied to embedded tumor detection.At present,in the embedded devices,it is difficult to directly use convolution neural network which has better performance but more parameters to detect the tumor.In order to solve this problem,this paper designs a low cost,low power consumption universal embedded terminal accelerating detection method which is based on the residual network,this method is suitable for the embedded terminal tumor detection.The experimental results in the stomach tumor data sets show that accelerating detection method greatly reduces the tumor embedded terminal consumption of memory and computing,and the accuracy of tumor detection is maintained at 94.89%.