Research On Image Classification Based On Deep Learning Algorithm

Image classification technology has been a hot topic in the field of computer vision. Deep learning algorithm establish hierarchical feature automatic extraction model and get close to image high-level semantic features, which has made a breakthrough in the field of image classification in recent years. Convolutional neural network (CNN) is a high recognition rate deep learning algorithm, which obtain invariant features of translation, scaling and rotation. However CNN model has a large number of parameters, which are trained with tens of thousands of labeled image. And with the increase of the layer number of CNN model, the gradient will gradually diffuse when the parameters are adjusted, which causes the parameters of bottom layers update slowly and difficult to achieve optimal, thus affecting the accuracy of image classification.In order to improve the broad applicability and feature recognition rate of CNN model in little labeled datasets, taking into account the advantages of restricted Boltzmann machine (RBM) unsupervise learn image features, we proposed a new method of image classification based on CNN and RBM mixed model R-CNN transfer learning. The method transfer pre-trained parameters in large labeled datasets to initialize CNN training model of little labeled target datasets, using RBM layers replace full-connect layers aims to unsupervise training RBM layers weight. Finally using little number of labeled images to supervise fine-tune the overall R-CNN mixed model. The structure and training methods of traditional CNN model are improved by our R-CNN model. The added RBM layers is not only fully connect all the feature maps to get rich invariant features, but also adequately unsupervise training to get high-order statistics features of target datasets itself. Therefore R-CNN model can improve the accuracy of image classification on little labeled datasets.Then to solve the parameters of CNN model bottom layers are difficult to achieve optimal problem, we proposed the improved convolutional coding (CE) algorithm to train every convolution layer of CNN model greedily. The CE algorithm valid convolutional encode input feature and full convolutional decode output feature, then the convolution filters are trained by minimizing the reconstruction error. The improved CE algorithm converge all convolution filter to optimal quickly with alternating iterative parameter optimization, which added the inter layer mapping relation, the weight regularization and the additional variables. Combined with the R-CNN model above, the ER-CNN mixed model based on the improved CE algorithm can be extended to a deeper level, and further improve the accuracy of CNN model image classification.The experimental results which verified on the MNIST, COIL-20 and CIFAR-10 datasets separately, show that R-CNN mixed model can be effectively applied in little labeled datasets, and improved CE algorithm can achieve better classification results when increase the layer number of CNN model.