Research Of Fine-grained Image Recognition And Classification Algorithm Based On Deep Convolutional Neural Network

With the rapid development of science and technology,in the daily life,the traditional information,which is based on texts and voice,is gradually replaced by the hypermedia information such as images and videos.Among them,the development results of regular image recognition and classification technology in the field of computer vision,pattern recognition,machine learning and deep learning have been widely used in real life.On this basis,combined with the actual needs,people are paying more and more attention to the task of identifying and classifying fine images,that is,the identification and classification of fine-grained images.Fine-grained image recognition and classification aims to detect hundreds of subcategories under a given base class.The difference from regular image recognition is that its purpose is to further identify sub-categories.The difficulty is that there are often only very small differences between subclasses.However,the features used by traditional image recognition classification methods are not enough to accurately describe the details of fine-grained images.At present,most of the methods for fine-grained image tasks apply traditional features to the local parts of objects to identify and classify the objects through specific parts,but this method is not universal.In recent years,the rapid development of deep convolutional neural network for multiple non-linear transformation of the image operation,extracted contains more abundant depth features,so fine-grained image recognition and classification tasks to see a new breakthrough.In this thesis we combine fine-grained image recognition and classification tasks with deep convolutional neural networks,and analyze the features of the depth convolution and the performance of different algorithms on the fine-grained images respectively from two aspects of strong supervision and weak supervision.1,Research of fine-grained image recognition and classification algorithm based on strong supervision.In addition to the regular category labels,such algorithms generally need additional manual annotation information to improve the accuracy of the algorithm during the model training phase.In the process of experiment,we first use the improved RCNN to train different detectors on the whole level of the object and on the part level of the object,and then add geometric constraints to the obtained area so as to select a useful detector.Finally,the whole and parts images are sent to convolutional neural network for training,equivalent to retain the global features and local features,which ensures the accuracy of the classification results.As a comparison,an improved solution to the training process of the detector of the algorithm component is proposed in the experiment.That is,the whole level detector of the object is first-trained,and then a part level detector is trained on the basis of the result so as to further improve the result.In order to train a better detector,the algorithm makes additional use of annotation information of objects in the dataset,so as to extract small blocks with discriminating ability and to facilitate object recognition and classification.2,Research of fine-grained image recognition and classification algorithm based on weak supervision.This kind of algorithm only needs regular category label in the model training phase without additional artificial information.The accuracy of the result is often improved by the improvement of the algorithm itself and the selection of the feature.Due to the above characteristics of the weak supervision algorithm,the original images were sent to different settings of the convolution depth networks during the experiment.Different networks are sensitive to specific features.Therefore,the resulting feature map can describe the original image from different perspectives.At the same time,we choose different ways to combine the obtained features to further enrich the descriptive ability of the features.Based on a two-stream model,this thesis merges the features extracted from two different networks in the same experimental process.Compared with the traditional features,this thesis finds that the extraction process of deep convolution features is a continuous abstraction process,and the ability to describe images is generally better than the traditional feature descriptors.In combination with the characteristics of fine-grained images,the discriminative regions are often the local parts of the image.Locating and detecting the local parts with discriminating ability is one of the difficulties in the fine-grained image recognition and classification tasks.For strong supervised algorithms,due to the use of a large number of annotation information,the part detectors can be trained by appropriate methods to obtain the desired partial part area,but it should be noted that additional annotation information requires a significant amount of upfront costs as well.For weak supervised algorithms,the process appears to be more similar to the traditional CNN training process,generally only requiring category labels,but also because it lacks additional annotation information,the requirements for representation of the image become higher and the features described need to be enriched.At the same time,it should also be recognized that weak supervision algorithms are the future trend of development.