Research On The Fine-grained Image Categorization Based On Convolution Neural Network

With the rapid development of science and technology,the theory of the deep learning pass through three phases:proposed,improved and leap development,deep learning algorithms has been successfully applied in many fields,such as text,sound,video,images,etc.Especially in the field of the image recognition application,image recognition technology caused the extensive concern of the academic circle,in recent years,image recognition technology based on deep learning has made great progress.Convolutional neural network algorithm is one of the most successful algorithms.Originated in the 1980s,convolution neural network algorithm effect is very effective in processing massive image data.The devel opment of hardware has provided convolution neural network development an unprecedented opportunity.Based on the research and analysis of existing convolution,this paper proposes an improved training model and application it on different image databases,the experimental results show that the proposed improved algorithm obtain the better recognition effect.The research and innovation points of the paper mainly as follows:Improve the training algorithm of convolutional neural network,improve the algorithm of convolutional neural network to separate the image according to the image features,and then get three probability values after training.The.K3 category table was obtained by experiment,using K3 category table and H judgment function recognition the different probability of each image and got the best classification result,experiments carried out on the Fine-grained image databases of the Ecuador moths and the Costa rica data.According to the data feature of the image segmentation it,division MNIST handwritten digital image up and down division form a simple structure.The algorithm is used in the CNN₁ designed?lenet-5 and AlexNet model.Due to the symmetry of the data of Ecuador and Costa rica moth,according to the vertical axis divides the pictures,the algorithm is validatiorned on CNN₂ designed and VGGNet model.This paper puts forward improvement methods to deal with the problem that over-fitting of the neural network to reduce the generalization performance of the training data,the article improve to a data set enhancement method,transformation the image makes the final amount of training data to ten times the raw input data.Analys is of experimental results,the result of the improvement method is higher than other methods,shows that the improved method is better.