Fine Grained Image Classification Based On Convolutional Neural Network

Fine-grained image classification refers to distinguishing different sub-categories from one large category,and has broad application prospects in commodity retrieval,vehicle identification,flower recognition and medical image recognition.Fine-grained image recognition is different from general universal object recognition.The inter-class difference is small,and the intra-class the difference is large.Discriminative features often local on subtle parts.Fine-grained classification is a very challenging task due to the diversity of foreground objects in terms of pose,scale and color.In this paper,based on the advanced algorithm in recent fine-grained image classification,the part localization algorithm and feature extraction technology of foreground objects are studied to improve the accuracy of fine-grained recognition.The main work of this paper is as follows:1.This paper studies and reproduces various advanced algorithms for fine-grained image recognition.They are local-based,feature-based,reinforcement-based,object-based and pose-based recognition algorithms.2.This paper makes a visual analysis of the deep convolution characteristics of the neural network,and makes a statistical analysis of the deep filter response.The neural network can automatically focus on the areas that have an important influence on the classification,and ignore the unrelated areas.The characteristics of neural networks.3.In this paper,a spectral clustering local localization algorithm based on feature similarity is proposed,which can effectively distinguish different parts of foreground objects without manual labeling.And achieved good classification performance on three standard data sets of fine-grained classification.4.In this paper,a random local localization algorithm based on probability model is proposed to determine the probability that they are selected according to the importance of different regions.The algorithm uses the probabilistic model to locate the locality without adding additional parameters,and achieves a good classification effect on the fine-grained dataset.