Convolutional Neural Network Based Transfer Learning Approaches With Its Application On Fine-grained Image Recognition

Image recognition technologies are widely used in the Internet industry,financial industry and security,which are mostly mature.However,there are many gaps in the study of fine-grained image recognition.One aspect is that current image recognition algorithms mainly focus on ?large? category recognition,such as cats and dogs,houses and cars.Little attention is paid to fine-grained recognition,for example,distinguishes Alaskan dogs and huskies in the same category as dogs.On the other hand,the process of collecting fine-grained data with labeled information usually requires expert knowledge in a specific field,which can be very time consuming and difficult,therefore,the field of fine-grained image recognition lacks sufficient and accurately labeled training samples.Due to the limited training samples and the subtle differences between subcategories and large intraclass differences,fine-grained image classification is a challenging research topic.The traditional fine-grained recognition algorithm is usually based on characteristics of artificial design,such as the local area location and the key location labeling points so as to obtain better classification accuracy.However this process needs enormous human and material resources.Therefore,it is necessary to construct a fine-grained image recognition model with high classification precision and strong generalization ability without using manual annotation information to make full use of the existing small number of marker samples.This paper focuses on the method of identifying fine-grained images.The main tasks are as follows:(1)Construct a fine-grained image database,such as Stanford Dogs,CUB 200-2011 Birds,and Oxford 102 flowers database from open datasets for image recognition and algorithm horizontal comparisons.(2)We research on the convolutional neural network model,convolution,pooling,full connectivity and activation functions in the convolutional network.We summarize the techniques for preventing overfitting in neural network models.Finally,several classical deep convolutional neural network models such as Alex Net,VGGNet,Google Net,Res Net,and Dense Net are analyzed and compared.(3)We proposal a convolutional networks based transfer learning algorithms framework.Firstly,the convolutional neural network is studied for the image feature representation capabilities.Then the transfer learning criterion based on the deep learning model are studied.Then the structure and knowledge of the pre-training model on Image Net large image dataset are transfered to the field of fine-grained image recognition.Those classic models(Alex Net,VGGNet,etc.)are used as feature extractors to obtain high-level feature representations of fine-grained images.Finally,a multi-layer perceptron model is constructed based on the extracted features to realize the recognition of fine-grained image data.(4)We study the influence of transfer learning of convolutional neural network models with different structures and different depths on the accuracy of fine-grained image recognition.In order to prevent overfitting,the weight decay and dropout operation are added to the multi-layer perceptron model to enhance the generalization ability of the model and improve the model prediction accuracy.We give the optimal prediction models for three fine-grained data sets and compared with previous studies.Our experimental results show that the feature representation based on deep convolutional neural network transfer learning have a very good discriminative ability.With a simple MLP as classifier,we achieve very good classification results on various fine-grained recognition task such as Stanford Dogs dataset,CUB 200-2011 Birds dataset,and Oxford 102 flowers dataset.On all datasets our methods outperform the state of arts and raising the recognition rates to 92.6%,74%,93.8% respectively.