Fine-grained Image Recognition Based On Transfer Learning

The computer vision classification research mainly aims at two types recognition tasks: the coarse-grained image recognition classifies objects with multiple main-classes(e.g.,classifying dogs,cats and fruits),while the fine-grained image recognition focuses on distinguishing different sub-classes under the main-class(e.g.,categorizing the sub-classes Husky and Samoyed of main-class dogs).Deep learning models build the deep features by combining shallow features,the semantic information of deep features is highly discriminative for classification tasks,which also greatly improve the performance of fine-grained image recognition.As is knowing,the deep learning model performance depends on a large amount of labeled data,but the labeled images are challenging for fine-grained tasks.On the one hand,it is difficult to collect samples with diverse categories;on the other hand,labeling fine-grained images requires the participation of domain experts,which is high-cost.This paper introduced transfer learning to transfer knowledge from the similar tasks to fine-grained image recognition tasks,to alleviate the limitation of insufficient datasets of fine-grained recognition tasks.This paper proposed two general models based on deep learning and transfer learning,which can alleviate the difficulty of obtaining and expanding fine-grained images,and improve the accuracy of recognition tasks.Besides,this paper adopted a regularization method which can be used in conjunction with the above two models for further improving the classification performance.The specific research work and innovations were as follows:(1)We proposed an unsupervised fine-grained image classification model based on transfer learning,the model was composed of a twin network that constructed by a two-branch neural network.The two-stage training strategy was used,in the first stage,the basic network was pre-trained by a supervised way so that the general features of fine-grained images can be quickly learned;in the second stage,datasets were expanded through the search engine,and the model was fine-tuned in an unsupervised manner for transferring knowledge.The module was a general data expansion and performance improvement method.Compare with the state-of-the-art models Bilinear CNN,WebData and L2 SP,the experiments showed that the twin network classification model had high accuracy,which can achieve 91.9% and 90.2% average accuracy on fine-grained datasets Stanford Car and Aircraft.(2)We proposed a bins similarity algorithm for fine-grained image classification model based on domain adaptation,in which bins similarity was adopted to measuring and selecting source domain datasets that similar to the target fine-grained datasets,while the selected source datasets were used to expand the target task datasets.The Image Net and i Nat were chosen as the source domain,multiple bins were divided in the feature space,and the similarity was calculated by statistics between the source and the target domain.The process of selecting similar domains was equivalent to expanding data for fine-grained image recognition tasks,while knowledge was transferred from the selected similar domain to the fine-grained recognition tasks.Compare with the state-of-the-art models EMD,SJFT and L2 SP,the experiments indicated that the bins similarity classification model can achieve 86.6% and 98.4%average accuracy on fine-grained datasets Stanford Dog and Oxford Flower.(3)We proposed adopting a regularization method based on the starting point that can be used in conjunction with the above two models.The method assumed the perfect solution should be near the starting point,then the effective search space of the loss function was constrained during the process of finding solutions in knowledge transfer,which improved the model performance.Compared with the original models,the experiments showed that the twin network with starting point regularization can improve 1.0% and 0.8% average accuracy on Stanford Car and Aircraft,while bins similarity classification model with starting point regularization can improve 0.2% and0.1% average accuracy on Stanford Dog and Oxford Flower.