Research Of Image Classification And Face Recognition Based On Deep Learning

Deep learning has an excellent performance in the field of computer vision.In recent years,convolutional neural networks(CNNs)based image classification algorithms have made many improvements in optimizing the network structure,preventing over-fitting,and accelerating the training speed.This paper first proposes training dataset optimization using a combination of active learning and deep learning,and then uses joint loss function supervision training to improve the performance of CNNs in image classification applications,and applies the improved joint loss monitoring algorithm to face recognition.Finally,the joint loss supervision method was extended to image super-resolution.Training CNNs often requires a large amount of labeled sample data,but it is not that the more samples are trained,the better.We have introduced active learning for sample set optimization to filter out more valuable samples.We improved the classical posterior probability sampling algorithm and proposed an active learning algorithm based on entropy ranking.Experiments show that active learning algorithms,especially our improved algorithms,can filter out more valuable samples and improve network performance.For the image classification task,we believe that if the homogenous image features are more aggregated and the heterogeneous ones are more dispersed in the feature space,the classification effect will be better.Therefore,we propose to use both mean square error(MSE)and cross-entropy loss function to supervise training at the same time.To optimize the MSE loss,we can aggregate the features of samples within the same class.Optimizing the cross-entropy loss can disperse the features of heterogeneous samples.For this reason,we also propose a simple and effective sample organization method for training.For face recognition task,the cosine distance metric is more convenient and effective than Euclidean distance.We replace the MSE loss in the above joint loss function with the cosine distance loss,and propose a simple cosine distance loss backward propagation method.The existing CNNs based image super-resolution algorithms often only optimizes the MSE between the reconstructed image and high-resolution label image at the pixel level in the training process.Based on this,we add a joint loss supervision method to cascade the image super-resolution network and the image classification network,and optimized the super-resolution network by optimizing the MSE between the high-level features of the reconstructed image and the high-resolution label image to obtain more high frequency details of images.Experiments show that our algorithm can reconstruct better subjective and objective quality.