Fine-grained Image Classification Based On Convolutional Neural Network

Fine-grained image classification,as the basic task of many computer vision application scenarios,has made great progress in recent years.With the continuous advancement of deep learning technology and the rapid increase of computing power,fine-grained image classification has or even exceeded the recognition ability of human experts in some scenarios.However,the fine-grained image classification task itself has the characteristics of large intra-class and small inter-class variance,as well as the lack of sufficient and effective training data and other practical factors,leading to a series of problems such as high dependence on the amount of training data,high feature dimensionality,and low efficiency in actual applications.In this work,in order to make the fine-grained image classification algorithm easier to deploy and apply,we conduct in-depth research and proposes solutions for the low performance data enhancement methods during training the fine-grained image classification task and the high computational cost of classic methods.The main contributions as follows:This work propose a joint classification algorithm of fine-grained image data enhancement and feature fusion based on attention mechanism.The algorithm uses the channel attention mechanism to extract high-information channel,based on which a dualthreshold segmentation algorithm is applyed to divide the input image into key areas,subkey areas and background areas.Through the cropping of key areas and the preservation of sub-key areas,a data enhancement algorithm based on area cropping and erasure is completed.The enhanced data can force the neural network to focus on more detailed areas of the object instead of only focusing on limited local information.At the same time,the fusion of the local features extracted by the attention mechanism and the global features can enable the neural network not only to obtain global semantic information,but also to focus on the discriminative features of objects.The experiment results show that our method has completed more effective data enhancement,and has better performance on the standard dataset than the classic methods.This work propose a fine-grained image classification algorithm based on multi-scale semantic bilinear features.Traditional bilinear pooling will increase the accuracy rate while bringing exponential growth in feature dimensions.The large computing cost makes it difficult for bilinear pooling to be deployed in practice.This work proposes a semantic grouping mapping module with multi-distance to obtain semantic bilinear features,which dimension is consistent with the ordinary neural networks.In view of the difference in resolution and semantic information of shallow and deep features,a semantic grouping method with multi-distance constraints is introduced,which makes the grouping results more reasonable.By extracting semantic bilinear features of different scales at different stages of the neural network,multi-scale semantic feature fusion is proposed,so that the fusion semantic information has complementary and enhanced effects.The extensive comparative experiments and ablation study validate that the proposed method has good performance on three standard datasets.The computational cost is also lower than that of the classic bilinear feature methods,and our algorithm is easier to deploy.