Fine-grained Image Classification

The era of big data and development of computer have seen the huge development of deep learning,semantic-level and instance-level classification have made great progress.Fine-grained image classification,which focuses on recognizing similar subordinates,e.g.different species of birds,has become a hot topic.Based on the collection of considerable data and the breakthrough of GPU computation,this paper tries to use intrinstic characteristics of fine-grained tasks to deal with the classification of fine-grained image classification with deep learning.Without bounding box or part annotation,this paper proposes a classification framework based on image parts.Two variants of the framework are part fusion model based on adaptive mechanism(PFNet)and part fusion model based on Gaussian mixture model(GMNet).Both of them have achieved good results in experiments.The main contributions of this paper are summarized as follows:(1)Based on low inter-class and high intra-class variances of fine-grained images,a new classification framework is proposed.Based on the framework,a part fusion model based on adaptive mechanism(PFNet)is proposed.It can fuse different image parts adaptively,including easy parts,hard parts and background parts.PFNet consists of a part feature extractor and a two-level classification network.The classification network consists of part-level and image-level loss functions.The former allocates different weights to different parts adptively,and the latter trains the image-level feature fused from part features.(2)The feature fusion method used in PFNet is relatively simple and cannot model the whole distribution.As a result,this paper furthe r proposes a part fusion model based on Gaussian mixture model(GMNet).GMNet incorporates a Gaussian mixture layer to model the distribution of image parts.It uses several Gaussian components to model their distribution and fuse them to an image feature.Training process consists of two loops.Outer loop is the optimization of the whole network,and the inner loop is about the EM algorithm used in Gaussian mixture layer.The two loops are combined and optimized by gradient propagation.(3)Classification results,parameter study,ablation study and qualitative analysis on PFNet and GMNet are given.Experiments demonstrate state-of-the-art or comparable performance on CUB-200-2011,Stanford Cars,FGVC-Aircraft and Stanford Dogs.