Research On Fine-grained Image Classification Based On Deep Learning

In recent years,with the rapid development of digital multimedia technology,the number of image data is increasing.How to effectively use computer vision technology to manage and classify image data has become a hot topic in the field of computer vision.Image classification is the basis for image management.Fine-grained image classification belongs to the category of image classification.However,because its data is very similar in appearance,at the same time,intra-class gaps are larger than inter-class gaps,making fine-grained image classification a more difficult and worthy research direction.For fine-grained image classification,we focus our research on digging the discriminative region in fine-grained images with weak supervision,and make full use of the calculated discriminative region images for classification.On the basis of the existing work,two discriminative force calculation methods,fixed size and adaptive size,were proposed using object saliency and spatial transformation respectively.In the fixed-size discriminative region calculation method,the weakly supervised fine-grained image classification method based on object saliency map,we do not use additional object location annotation information.We use the attention mechanism of the model to propose a new saliency map calculation method to calculate the fine saliency value on an object,and then calculate the image by combining the receptive field and saliency map of the convolutional neural network.In multi-scale discriminative regions,the discriminative multi-scale region is finally used to train multi-scale models,and by combining different-scale classification models,the optimal scale combinations are fully utilized by using the object-level image information and the local details of the object.The experimental results on the CUB200-2011 dataset show that the proposed method can find the discriminative area in the image and improve the image classification accuracy.In the adaptive size discrimination region calculation method,namely the adaptive supervised spatial transform based weakly supervised fine-grained image classification method,a spatial transformation layer is introduced into the convolutional neural network,and each image space is automatically learned by the training model.Transforming parameters,performing image transformation on the input image,removing interference from unrelated backgrounds,obtaining discriminative regions in the image,and finally using these calculated image patches to train and test the model,and also introducing a multi-scale model framework,Combining object-level and local-level models together to train the convolutional neural network can make full use of the object-level image feature information and the image feature information of the local level of the object.Finally,the cross-entropy loss function and the Rank loss function are used.The model was alternately trained.The results on the CUB200-2011 dataset show that this method effectively improves the classification performance of the model.