| As one of the most important research directions of computer vision,fine-grained image classification aims to identify which subcategory the image specifically belongs to.For fine-grained image classification,the differences between the target objects of different subcategories are subtle,while the differences between the target objects of the same subcategory are obvious due to illumination,posture,angle,etc.Therefore,fine-grained image classification remains very challenging.To address the above challenges,learning subtle interclass differences is the key to solve fine-grained image classification.At present,the mainstream fine-grained image classification algorithms usually locate discriminative object parts,and then learn rich feature information through feature extraction of these discriminative object parts.Although many fine-grained image classification algorithms have made much progress,there are still three urgent problems to be solved.Firstly,many existing methods tend to focus only on the most discriminative region of an object but ignore the second discriminative region.Secondly,some methods focus only on detailed parts but ignore the surrounding context information of each detailed part.Finally,most existing methods generate discriminative regions by data augmenting or extending layers of the network,but there is no guarantee that these regions are highly discriminative.To solve the above problems,this thesis proposes a weakly supervised sampling network for discriminative regions(SNDR).Firstly,a location image that can magnify object is generated by target object locator.Secondly,some evaluation metrics are used to measure amount of information and discrimination of region proposals.Finally,a discriminative region sampler is proposed,which can generate the most discriminative image and the second discriminative image(complementary image)with the help of the evaluation metrics.Experiments demonstrate that SNDR performs well on three benchmark datasets,such as CUB-200-2011,Stanford Cars and FGVC-Aircraft.However,as the vast majority of current algorithms,SNDR also cannot avoid sampling irrelevant noise regions that damage the network learning.In addition,there are few methods to optimize the extracted irrelevant noise regions.More often,many algorithms allow the network to extract features from these irrelevant noise regions.To solve the problems existing in SNDR,this thesis proposes a weakly supervised sampling optimization network for discriminative regions(SONDR).Based on the SNDR model,a target object relocator is firstly proposed to generate relocalized images.Secondly,before sampling discriminative regions,an edge noise remover is used to reduce the interference of irrelevant noise to sampling operation.Finally,a discriminative region optimizer optimizes the most discriminative image and complementary image.Experimental results demonstrate that SONDR model achieves excellent classification performance.To sum up,the main contributions of this thesis are summarized as follows:(1)In order to learn the global features of target objects better,we propose a target object locator and a target object relocator.The two kinds of object images generated by them both play the role of data augmenting.Among them,the target object relocator can generate an enlarged and undistorted relocation image,which is beneficial to the network to learn original global features of a target object better.(2)Aiming at the problem of the instability of discriminative regions,we propose a discriminative region sampler and three evaluation metrics.The three evaluation metrics are respectively the ratio of total activation values in a region proposal to total activation values in the connected region,the ratio of object area in a region proposal to region proposal area and the ratio of object area in a region proposal to object area in the connected region.Under the guidance of these evaluation metrics,the discriminative region sampler can sample the most discriminative region and complementary region with high quality,which not only magnifies each detail part,but also does not lose the surrounding context information of each detail part.(3)In order to reduce the harmful influence of irrelevant noise regions on fine-grained feature learning,we propose an edge noise remover and a discriminative region optimizer respectively.Among them,the former can reduce the computation amount and the interference of irrelevant noise to the discriminative region sampler.The latter can remove the irrelevant noise regions which is not conducive to network feature learning to some extent. |
