Research On The Theories And Methods Of Designing Convolutional Neural Network Sub-structures For Image Classification

In the field of computer vision,Image classification is a fundamental task and re-search hotspot.It is an important part of many vision processing tasks,such as:object de-tection,semantic segmentation,and object tracking.Therefore,the study of image classi-fication technology has important theoretical significance and application value.In recent years,with the advent of the big data era,the improvement of computing power and the de-velopment of deep learning technology,a variety of image classification algorithms based on convolutional neural network(CNN)structure design have been proposed,which have greatly promoted the development of image classification.However,the CNN structure has characteristics such as complexity,diversity,and variability in connection methods.These factors lead to the challenges for image classification in important issues such as network structure construction and network model light-weighting.It is urgent to research and design efficient CNN structure in the field of computer vision.Therefore,this paper has launched the research on the design of sub-structure of CNN for image classification.This dissertation aims to improve the feature extraction performance of the CNN for image classification.In view of many characteristics of the substructure of CNN,the structure of CNN is studied from two aspects:basic feature extraction structure design and feature enhancement structure design.At the same time,the use of network structure design to solve the transfer of knowledge structure in transfer learning and the problem of noise label in image classification are discussed.The specific research content and main innovations of this dissertation can be summarized as follows:(1)Since the problem that the existing activation function cannot adapt to the output change due to the limitation of the function form,from the perspective of feature extraction structure design,a parameterized deformable exponential activation function method is proposed.This method firstly introduces a deformable exponential function form with excellent function properties for the activation function,and at the same time introduces a learnable parameterization factor to adjust the response scale of the activation function to the output of different network layers.Finally,this paper proposes a weights initialization method which is suitable for the activation function.The activation function effectively improves the ability of nonlinear modeling for the CNN.(2)Since the problem that single-loss supervision in the existing CNN cannot ef-fectively construct the feature distribution,this dissertation studied the supervision infor-mation in the network training process and proposed a Hybrid supervision loss function method for image.The method first introduces two auxiliary loss functions to constrain the compactness within each feature cluster and the separation between feature clusters.Then,the paper proposed to use multiple losses to supervise a single classifier to reduce the difference between the extracted features.This paper theoretically proves the superiority of the hybrid supervision loss function.This method can improve the feature extraction capabilities of CNN and can be extended to text classification.(3)Since the problem that the existing attention module cannot effectively adjust the feature enhancement methods for different network layers,from the perspective of fea-ture enhancement structure design,an attention module based on adaptive adjustment is proposed.This method first integrates the existing attention mechanism,and then intro-duces a set of learnable parameters to weight the attention sub-modules,and learns the adaptive distribution of attention modules in different network layers,thereby improving the feature enhancement capability of the network.(4)Since the problem of low discrimination between samples due to the same pro-cessing of samples in the training phase,this paper carried out a research on structural design based on feature enhancement between samples,and proposed an attention mod-ule based on batch dimensions.This method first uses the in-sample attention mechanism to generate multi-dimensional attention weights.Then it use attention weights to generate sample importance and perform normalization operations between samples to generate rel-ative importance between samples.Finally,the relative importance of the sample is used to weight the sample features as a whole,which effectively improves the discrimination between sample features.(5)Since the problem that the information of the teacher model cannot be well trans-ferred to the student model in the existing transfer methods,this paper proposed an al-gorithm based on multi-group knowledge transfer.This method first introduces a multi-teacher model for joint transfer,and then simultaneously migrates the final prediction layer results and middle-level features of the teacher model,and quantifies the middle-level transfer features with group to eliminate the problem of information redundancy.Finally,the similarity measurement of features between samples is constructed to effec-tively improve the effect of knowledge transfer.(6)Since the problem of noise in labels under the existing supervised learning frame-work,This paper proposed a noise label classification algorithm based on sample feature re-calibration.Th is method firstly augments the samples in the input spaca with fusion,and performs implicit linear regularization on the networkto simplify the complexity of model training.Then it use the fusion attention mechanism to weight the samples in the feature space,so as to achieve the affect of sample feature enhancement.Finally,It in-troduced the label softening method in the label space to reduce the influence of noise labels on the supervision signal,so as to explore the implicit information of the label.This method effectively improves the anti-noise ability of the CNN.