The Research Of Computation Efficient Convolutional Neural Network Classification Algorithm For Mobile Devices

Deep learning has become a popular research area since AlexNet made a great success in ImageNet 2012.The researchers use deep learning solve the problems in computer vision areas such as image classification,semantic segmentation and object detection and make a great progress.There are many excellent CNN models are proposed in recent years.However,most of these CNN models have a heavy computation and the size is very large,so these models can only be used in server and run on professional GPU.But smartphone and FPGA devices have a great demand of deep learning.So how to design computation efficient CNN models which can run on the mobile devices become a big problem.And I select the computation efficient network design as my research topic.Currently,the main computation efficient CNN model is MobileNet,the results of the model are not good enough.I propose a multi-level residual network unit,which use depthwise separable convolutions and group convolutions.The unit reduces the computation by multi-level residual connection.Then I propose a computation efficient convolutional neural network classification algorithm for mobile devices which is called EENet,EENet decreases the computation of the model without affecting the accuracy by multi-level residual network unit.I test the EENet model in Food101,ImageNet and Place365 dataset,the experiment shows that EENet result is better than SqueezeNet.And the computation of EENet is only 43 M FLOPs.I also propose a self-adaptive incremental dynamic quantization algorithm,which can quantize a majority of CNN models.The proposed algorithm can quantize the model to low bits incrementally and without affecting the accuracy.The experiment shows the proposed algorithm can quantize ResNet model and the results are better than TWN and other algorithm.The proposed algorithm can quantize the EENet to 6 bits and the accuracy decrease less than 1%,the size of quantized EENet is only 0.53 M and the computation is 43 M FLOPs,The quantized EENet can run on smartphone and other mobile devices.