The Research On Fine-Grained Object Classification Algorithm Based On Depth Learning

Image classification is a very active and classical research field in computer vision.With the development of computer vision technology,the traditional image classification is difficult to adapt to the growing demand for retrieval of people's life,and the classification of fine-grained image gets more and more s concern.Fine-grained classification is a sub-domain of object recognition,specifically a task that identifies a category or subcategory under the same category(eg,birds or fish).Because the traditional algorithm relies on a large number of artificial annotation information,there are great limitations in dealing with the natural data of the original form,and in the face of the fine-grained classification problem,due to the lack of information on the capture object structure and lack of detail The description of the area makes it difficult to improve the classification accuracy.With the development of deep learning and computer vision in recent years,a large number of deep learning algorithms have been proposed.Deep convolution neural network has created new opportunities for fine granularity classification,and it has also promoted the great development in this field.The current image classification generally refers to the classification of coarse-grained object categories,such as:vehicle,animal,food and other different categories of differences between the comparison,and the lack of research on similar objects between the fine,it is difficult to fine-grained image comparison Analysis,can not do the exact distinction and multi-level classification.Therefore,it is of great significance and practical value to use fine-grained objects to do fine,accurate and efficient classification,and to generate information that users can understand and apply.The main research work is as follows:(1)This paper analyzes the difficult and present situation of the existing fine-grained classification algorithm.The main difficulty lies in the large intra-class difference and the small inter-class difference in the image,the commonly used target feature descriptor,the calculation principle of the commonly used feature,This paper summarizes the advantages and disadvantages of the algorithm with excellent performance in the commonly used feature extraction method,and selects the feature extraction method suitable for this paper.(2)In order to capture the details of the information area,to achieve the object area detection,the use can reduce the over-fit convolution neural network model,through the weak supervision of the way to automatically learn the characteristics of sample data,do not rely on component labeling information,only based on artificial experience to design features extractor,the output function is a nonlinear line segment adjustment functioin,and the fifth layer convolution layer feature is used as the image feature to experiment.(3)The method of multi-scale convolution feature matching is proposed to construct the image feature representation from the local area feature generated in the convolution feature.The SVM classifier is used to classify the fine-grained image.Through a lot of experiments,it is proved that the algorithm has good classification effect,Leading the current traditional classification algorithm.