Application Research Of Fine-grained Image Classification Based On Deep Convolutional Neural Network

With the unstoppable development of correlative technologies in the field of computer vision,in daily production and daily life,people have asked more precisely demands of image classification.Therefore,the finely grained image classification has become a problem for many researchers to be increasingly concerned about today.In the coarse-grained image classification task,the goal of image classification usually is the category classification in a traditional sense of "cat","dog","bird".The finely grained image classification's target is the different subclass categories under the coarse-grained category,such as classifying different species of birds,"Arctic Tern" and "Caspian Tern",etc.Compared with the coarse-grained image classification problem,the classifying difficulty of the finely grained image is the great difference in the intra-class feature,while the inter-class feature differences are small,and the complex background feature in images also create interference to classification.In this paper,how to effectively extract the refined image features in fine-grained images is studied.The method of image feature extraction with hierarchical refinement is applied to the classification of fine-grained images,and different classification and recognition models are constructed for different data sets.The main works was as follows:(1)Based on the multistage features,constructed a finely grained image classification model of birds.Because the originally designed intention of the convolutional neural network model is for the coarse-grained images' classifying work,so its classification effect is unsatisfactory in finely grained image classification work.As for the difficulty of the finely grained image classification work,this paper proposes a method of hierarchical feature extraction,which divides the finely grained image feature extraction of birds into original image class feature,birds object class feature,birds parts class feature and refined birds image in layers.and combined with three classes of feature Map input classifier in order to classification.The experimental result of the CUB200-2011 data set has showed that the hierarchical feature extraction method can refine the process of image feature extraction,capture the finely grained feature in image as well as improve the classification accuracy of finely grained image.(2)Based on double convolutional neural network,constructed a plant leaf recognition model.In the process of researching the classical convolutional neural network model to identify the leaf image types,it was found that the misidentified leaves have similar edge shapes as the target blades.We can obtain a characteristic mapping visual figure in the pixel space of the original image by deconvolution reconstructed the characteristic mapping leaf image from the extraction of the convolution layer.As we can see from the analysis of the visualization that the convolution kernel is more responsive to the blade edge shape than the other blade features,so there is a defect that the similar edge shape interferes with the blade recognition.As for this problem,this paper proposes a method of double convolutional neural network model and refining the leaf feature extraction.The two paths respectively input the original leaf image,the corresponding texture image block,extract the edge shape features and texture features of blades,and also fuse the two feature input classifiers to classification.The experimental results from the Flavia data set has showed that the texture features can avoid the interference of similar edge shape features and improve the accuracy of blade recognition.In this paper,both models applied the idea of refinement image feature mapping extraction,expanded from the classic cub200-2011 data set to the Flavia leaf data set in the forestry field.Different depth convolution neural network model structures are constructed for different fine-grained image samples.Experimental results show that this method can effectively capture the fine features in the fine-grained image and improve the classification and recognition effect.