Deep Neural Network Model And Algorithm For Image Classification

Image classification is a core scientific problem in the field of computer vision,and it has always been a research hotspot.With the development of deep learning technology,the focus of image classification research has shifted from traditional machine learning methods to deep neural network modeling,achieved better accuracy.However,this increase in accuracy relies on a large number of accurate training samples.In practical applications,it is usually difficult to obtain so much accurate training data,because a lot of data contains noise,and even data with label noise.Therefore,the trained classification model is easy to overfit the noisy data,resulting in poor classification effect.It can be seen that how to improve the generalization ability of existing image classification models on data containing noise is still an urgent challenge in the field of image classification.In response to the above findings and considering the importance of the features extracted by the network for image recognition,this paper designs two noise-tolerant deep neural network algorithms for image classification.The specific methods are as follows:(1)A self-paced learning based hybrid dilated convolutional neural network model for image classification is proposed.In order to alleviate the damage of feature information caused by multiple pooling in deep convolutional neural networks and ensure larger receptive field,this paper uses a hybrid dilated convolutional network structure to replace the conventional combination of convolution and pooling,and improves the network part.Further,in order to alleviate the impact of noise samples in the training set on the classification effect,the introduction of self-paced learning to improve model training,reduces the impact of image noise samples to improve the robustness of the classification model.(2)A self-adaptive image classification algorithm tolerant to label noise is proposed,which makes full use of the characteristics of meta-learning and channel attention-wide residual neural network.First,on the basis of constructing a deep wide residual neural network,the channel attention mechanism is introduced to rebuild each wide residual module,and the contribution of each feature channel to the model is obtained and the feature channel is weighted by the contribution.Then,the meta-weight-net is introduced and the weights of training samples are adaptively learned from the data under the guidance of metadata,and adaptive weighted samples are used to improve the tolerance of the network model to noise.Finally,the Hard Bootstrapping Loss is introduced into the meta-network model,which alleviates the tendency of the cross-entropy loss function in the meta-network model to noise data,and realizes the further improvement of the generalization ability of the classification model on the label noise data.In order to prove the effectiveness of the model in the paper,experiments were performed on each module of the proposed method on multiple datasets,and a comprehensive experimental analysis of the proposed method was performed,and better results were obtained.