Research On Hierarchical Loss Function Of Lightweight Convolutional Neural Network For Image Classification

Image classification is one of the four key technologies in the field of computer vision and the basis of the other three technologies.In the image classification task,how to extract more comprehensive and abstract features of the image is the key to solving the problem.Lightweight convolutional neural network(LCNN),as a convolutional neural network(CNN)with fewer parameters and lower computational complexity,can autonomously extract the feature expression of data,and has been widely used in image classification tasks.The loss function is an important part of the LCNN model.It is used to evaluate the difference between the model prediction results and the real value,and determines the update of the model parameters.It is an important criterion to measure the feature extraction ability of the model for the original data image.Therefore,the research on the loss function based on the LCNN model is of great significance in improving the performance of the model.However,the current research on loss functions mainly focuses on the design of a single loss function,aiming at optimizing the intra-class and inter-class distances,and lacks the discussion and analysis on how to obtain representative class prototypes to regularize intra-class distances.Secondly,the existing work mainly analyzes the output characteristics of the feature expression of the penultimate layer of CNN,and lacks the discussion on the feature expression of the middle layer of CNN.Thirdly,the existing work on the improvement of the loss function is mainly aimed at tasks such as face recognition and image retrieval.There are fundamental differences in the discriminant criteria on which different tasks are based,and the performance gains are limited when the corresponding loss functions are applied to image classification tasks.In view of the deficiencies in the above research,this thesis focuses on the LCNN-based image classification task and conducts research on the optimization method of the loss function.The main work is as follows:(1)To obtain more representative class prototypes,this thesis first proposes a polynomial kernel loss function.By using the polynomial kernel function instead of the inner product in the traditional softmax loss function,the feature vectors that are difficult to separate in the low-dimensional space are mapped to the high-dimensional space,and more representative class prototypes are obtained to regularize the intra-class distance,which is more convenient for calculation.The distance between a sample feature vector and its corresponding class prototype.Secondly,a hierarchical prototype polynomial kernel loss function is constructed by using the polynomial kernel loss function,which alleviates the phenomenon of gradient disappearance and explosion during backpropagation,and improves the generalization ability of the LCNN model.(2)The output characteristics of different convolution units of LCNN are explored and analyzed,and a hierarchical loss function optimization framework is proposed.By fusing the knowledge learned by multiple loss functions in different convolutional units of LCNN,a comprehensive and abstract feature expression of the target object is obtained,which improves the classification performance of the LCNN model.Aiming at the collaborative optimization problem of hierarchical loss function,a hierarchical adaptive loss function optimization framework(HALOF)based on quantum genetic algorithm is proposed to construct a model for adaptive learning tasks.Based on the structural characteristics of HALOF,a two-stage retraining strategy is proposed,which improves the training efficiency of HALOF and improves the anti-overfitting ability of the model.(3)In order to further verify the effectiveness of the algorithm proposed in this thesis,experiments were conducted on multiple different LCNN models and different types of image classification datasets.Experimental results show that,compared with the baseline model,the proposed two algorithms improve the accuracy by about 10%on fine-grained image datasets.The article has 32 figures,15 tables and 91 references.