Image Classification Study Based On Convolutional Neural Network

Convolution neural network has become a research hotspot in the field of computer vision with its superior performance in image classification.In this paper,some improved methods are proposed based on the VGG network to reduce its overfitting problem and improve its image classification accuracy in complex scenes.Furthermore,an improved weight parameter optimization algorithm is proposed to accelerate the convergence rate optimization process,and improve optimization ability.The main work of this paper includes the following aspects.First,research background and significance of convolution neural network are introduced.Research status and progress of convolution neural network and weight parameter optimization algorithm are summarized.Some basic concepts related to the paper are introduced.In real world,due to the cluttered background,ambient occlusion,lighting changes and other factors,image classification tasks are very difficult.One purpose of this paper is to improve the accuracy of image classification tasks in complex scenes.Classification method of this paper is based on VGG network and it improves the latter in two points.The first point is to optimize model's structure and reduce model's over-fitting problem;the second one is to add into some target detection element to obtain a blend model with double loss function,where one loss function's purpose for optimizing parameters is to ouput the target object's bounding box coordinate point values,box length and width,and the other's purpose is to output the accurate classification of the target object.Since these two loss functions share weight parameters of the feature extraction layers,the feature map extracted from the model tend to locate on the target object region.As a result,feature extraction is focused on the area of target object,thus effectively excludes the background interference.Finally,improved model,original VGG model and GoogleNet model are compared by using commonly used data sets to verify effectiveness of the improved model.With the developing depth and complexity of convolutional neural network,more advanced optimization strategy need to be applied to handle it.An improved weight optimization algorithm based on Adam is proposed to reduce training time and improve performance of parameter optimization from two aspects: optimization speed of the original Adam algorithm near saddle point is slow,so an improved zero order optimization method is added to improve it.The main content is to track zero order information of the loss function in optimization process and add them into the original Adam algorithm as a feedback term.For the model's overfitting problem in the end of optimization process,a cyclic learning rate annealing method is proposed,which can effectively reduce overfitting.The improved algorithm is analyzed on the commonly used data set,which proves its superiority.Finally,the paper elaborates programming realization of the improved convolution neural network and design work of experiment based on concrete data set,and prospects future research work.