Fine-grained Image Classification Based On Convolutional Neural Network

Aiming at the classification bottleneck caused by fine-grained images due to high intraclass variances,low inter-class variances,complex backgrounds and posture differences,this paper designs a series of end-to-end weakly supervised learning classification models based on convolutional neural networks.Realizing the improvement of fine-grained image classification accuracy,the main research focus of this paper is how to accurately extract and identify finegrained features which are easy to be confused and difficult to be classified,and effectively reduce classification errors.The specific research content is as follows:1.Fine-grained Image Classification Based on Self-attention Scale-transformation Network.In order to improve the expressiveness of the network,we proposes a self-attention fusion module and multi-scale transformation to improve ability of extracting feature.In the attention-fusion module,we compressed the information the two dimensions of space and channel on the basic features extracted by ResNet50 to find the key features of each part and enhanced the parts features to make full use of attention maps that emphasizes detailed features.In the multi-size transformation module,the generation of the object-level image and the partlevel image map combines the semantic features and detail information of object in fine-grained images to solve the classification problem caused by the small difference between the finegrained classes.2.Fine-Grained Image Classification Based on Multi-branch Attention-Augmentation.The pre-trained Inception-V3 network is used to extract basic feature.In order to solve the problem that features are extracted from one part of an object and encourage the network to pay more attention to the discriminative features of different parts,we apply self-constrained attention-wised cropping and self-constrained attention-wised erasing on the central parts of the original images.It also improves the detection accuracy of object locations.Meanwhile,a central regularization loss function is proposed to constrain attention-augmented training process to obtain better attention regions and expand the gap between different classes of images.Com-prehensive experiments in three benchmark datasets show that our approach surpasses the state-of-art works.3.Fine-Grained Image Classification via Multi-stage Attention-Fused Convolutional Network.Fine-grained image classification,is becoming a more difficult and challenging field of computer vision because of its high intra-class variances and low inter-class variances on data content.For fine-grained image classification,our research focuses on using weakly supervised learning to extract discriminative features of fine-grained images to accurately classify the target images.Based on the existing work,a new network named Multi-stage Attention-Fused Convolution Network is proposed with supplementary data augmentation.The multi-stage attention-fused network consists of an attention-fused module based on InceptionV3 and an adaptive bilinear feature fusion block,while supplementary data augmentation is taken by the attention cropping and attention erasing to eliminate irrelevant background information.Meanwhile,part feature loss is added to further improve object positioning and classification performance of the entire network.Comprehensive experiments in three benchmark datasets show that our approach surpasses the state-of-the-art works,which validates its effectiveness in fine-grained image classification.In summary,all methods in this paper are oriented by the attention mechanism,using feature fusion and data augmentation to identify the semantic features of fine-grained images,and correcting attention areas of networks through the feedback of the additional loss function,which effectively improves the fine-grained Image classification accuracy.Comprehensive experiments in three benchmark datasets show that our approach surpasses the state-of-art works.