Rare Bird Sparse Recognition Via Component-Level Multi-feature Fusion

The tracking of rare birds has always been one of the focuses in the field of animal protection.With the popularization and development of monitoring equipment and technologies,it has became a research hotspot to implement intelligent bird observation unattended through out the day in the wild environment.However,at present,the main stream method for intelligent identification of birds in the academic community is based on the idea of deep learning,and most of these algorithms require large amounts of data for observation targets,which is in contrast to the reality of rare amounts of data for rare birds.In response to this phenomenon,a sparse bird sparse recognition method based on component-level multi-feature fusion was proposed.Taking into account the geometrical complexity of the birds image target in the key areas such as the head and body,different from the traditional processing method to complete the bird image as the object to be processed,the text adopts the method of obtaining the bird parts from the image and then identify the birds.The main work of this article is as follows:(1)Researched the commonly used algorithms for fine bird image recognition,and introduced the current research status of component detection,feature extraction and fusion,and classifier design.In the component inspection part,the typical part detection algorithm is summarized in detail.In the feature fusion and extraction part,the three levels of feature fusion are introduced.In the sparse classifier part,the typical sparse solution algorithm is introduced and compared.The merits of the algorithm.(2)In-depth understanding of the development of bird image fine recognition,and combing the development process of fine recognition,enumerating the algorithms that have achieved good results in the field of bird image fine recognition so far.And by summarizing the existing methods,the precise positioning of the component information plays an important role in improving the recognition rate.Based on this conclusion,the method clarifies the existing bird image component detection algorithm in detail,selects the bird component acquisition method based on the limitations of the existing algorithm and the small amount of rare bird image data,and makes a series of comparative experiments.According to the experimental results,the effect of component detection on bird image recognition is analyzed.(3)Based on the information of bird parts,this paper proposes a bird image component-level multifeature fusion algorithm based on feature fusion theory to solve the problem of bird image component information representation.In this paper,the typical method of feature fusion is studied firstly,and the feature-level fusion is selected as the feature fusion method according to the characteristics of the topic.The fusion features are color,global and local gradient features.The feature fusion algorithm proposed in this paper first extracts the components and The three characteristics of the global image,and the three features are normalized,and finally the features are merged by means of concatenation.In this paper,the sparse representation classifier is used to classify the final bird image.The feature of the bird image after the feature fusion is used as the sparsely expressed atom,the image of each type of bird is the most a class dictionary,and finally the test image feature is obtained.The sparse coefficients under the dictionary formed by the training set and the category of the birds images to be classified are determined according to the properties of the sparse coefficients.Finally,this article takes the CUB-200-2011 bird dataset as the main experimental object.The results show that for the rare bird recognition with small data characteristics,the text method has advantages in the recognition rate.