Research On Fine-grained Image Recognition And Segmentation Based On Deep Learning

In recent years,the number of image data is growing exponentially with the development of digital media technology.How to effectively use computer vision technology for image classification and image segmentation has become an important research branch in the field of computer vision.Traditional image classification is to identify coarse-grained categories,such as using computer vision technology to distinguish meta-categories: "dog","grape" and "bird".In many practical applications,images need to be classified at a fine-grained level,for example,to distinguish which subcategory of the "Grape" category the image belongs to: "Summer Black","Hong Ti","Sunshine Rose",etc.Fine-grained image classification is a popular research direction in the field of computer vision for such problems,and has wide application value in real scenes.The fine-grained image classification is a challenging study due to its small intra-class differences and large intra-class differences.However,the fine-grained image classification only obtains the category information.Usually,the boundary and attribute characteristics of the object need to be identified for further research,and the target needs to be segmented from the image.This paper explores the research of fine-grained image recognition and segmentation based on deep learning.The main works are summarized as follows: 1.Grape dataset(Vitis-15)is established: In this paper,a small and fine-grained image dataset is established.This dataset was manually captured in a natural scene of fruit and vegetable greenhouses around Wuhan and Yunnan.There are 6389 images and 15 grape subcategories.Due to the different shooting times,years and locations,it has the characteristics of large intra-class differences and small inter-class differences.2.Fine-grained image recognition based on multi-scale data fusion: In real-life scenarios,dataset generally suffers from the problem of imbalance.In order to solve this problem,this paper proposes a multi-scale data fusion method.Multi-scale image import can extract more discriminative information for image classification tasks,and the data amplification method is beneficial to solve the imbalance of different classes,so that the classification network can learn the characteristics of each category.For the Vitis-15 dataset,transfer learning and convolutional neural networks are used for classification.Based on the Alex Net,a multi-scale enhanced Alex Net(MS?EAlex Net)is proposed.The experimental results show that MS?EAlex Net is better than the current mainstream classification model on the Vitis-15.3.Fine-grained image recognition based on multi-scale image destruction and reconstruction: Extend the research of Vitis-15 fine-grained image classification task to CUB200-2011 fine-grained data set with large number of categories and complexityt,it is difficult to achieve better recognition accuracy by only using the classification network.In the past,most of the fine-grained image recognition was localized through annotations,but through the "destruction and construction" network,part discriminative information can be automatically obtained.This paper proposes multi-scale inputing(MS?DCL)and multi-scale detail enhancement(MSDB?DCL),So that the network can learn the discriminative information at the object and part level.Experimental results show that MS?DCL has higher accuracy than other existing fine-grained image classification methods.4.Fine-grained image segmentation based on Mask Grab Cut: In order to obtain the foreground part of the fine-grained image for further research and analysis,in view of the limitations of the traditional image segmentation algorithm and the insufficient boundary reservation based on the deep learning segmentation network,a combination of the traditional image segmentation method and the deep learning segmentation network is proposed Mask Grab Cut,this method combines the idea of cutout boundary in Grab Cut segmentation algorithm and the universality of Mask R-CNN semantic segmentation network.The experimental results show that the Mask Grab Cut method performs better than these two methods on CUB200-2011 dataset.