Image Classification Based On Feature Encoding And Deep Convolutional Neural Network

Image classification is a fundamental problem in computer vision.With the rapid development of Artificial Intelligence and computer vision,more and more universities and enterprises have invested a lot of energy into Image classification.Image classification,as the name implies,is to extract image features using the method of image processing and Artificial Intelligence,and then determine the classification of the image.In the traditional image classification algorithm,we first extract color features,texture features,shape features and spatial relationship features of the image,then train a classifier to classify the images.The classification accuracy of traditional image classification algorithms is limited by the typicality and distinction of features.In this paper,feature encoding and Multi-level Spatial Feature Pyramid are used to obtain more comprehensive and discriminative features.The main work of this paper is as follows:1.An image classification method based on deep network and gaussian aggregated encoding is proposed.Firstly,the features are extracted by the deep convolutional neural network.Then,we use Gaussian aggregated encoding to re-encode image features.Finally,the encoded features are input to the fully connected layer to classify the images.This method combines deep learning with Gaussian aggregated encoding,which makes features have richer semantic information,certain sparsity and higher classification accuracy.2.An image classification method based on deep network and multi-level spatial feature pyramid is proposed.The lower-level features of the network mainly represent the basic semantic information,and the higher-level features are more complex semantics information which is more discriminative for image classification.The lower-level features and higher-level features are effectively fused by multi-layer spatial feature pyramid,and then the fused features are input to the full connected layer to classify the images.It makes the feature representation of image information more comprehensive and specific.On this basis,the network model for image classification training can achieve higher classification accuracy.3.An image classification method based on Open CL parallel acceleration is proposed.Firstly,a lightweight convolution neural network model for real-time classification is constructed.Then,each layer of the convolution neural network is optimized by instruction vectorization.And parallel acceleration design is carried out for each layer of the network.The method is based on the Open Computing Language(Open CL),which can be applied to various heterogeneous systems such as CPU,GPU,and FPGA.The experimental results show that the image classification efficiency can reach 229 images per second on GPU(GTX 1080)platform,and 25 images per second on Stratix 10 platform.And the image classification accuracy is 93%.