Fine-Grained Object Detection Based On Deep Learning

Objects are one of the basic elements that make up a scene.It is important to understand that objects such as cars,airplanes,ships,pedestrians and pets in a picture are the scene to be judged and the contents of a picture to be understood.The changes of illumination,occlusion and deformation of different targets in the picture are very complicated,so that the related research of target detection has been a basic and difficult point of image processing.With the development of research on the technology of target detection,as a derivative problem of target detection,the research and application of fine-grained target detection is becoming a hot spot.Compared with general target detection tasks,fine-grained targets are more likely to have large intra-class differences and small inter-class differences,which makes the detection of fine-grained targets more difficult.So many methods can only classify the picture,not to detect objects.Many kinds of object in a picture is a normal condition,so does detection system is more practical.This paper studies the detection and identification of fine-grained objects in the picture,extracts the location of the specific object in the picture and the information of the corresponding category.First of all,the convolution neural network is mainly studied,its basic theory is introduced,and some important components needed for the structure and target detection of the two classical network models are introduced.Then,the development of target detection is reviewed and compared to compare the advantages and disadvantages of different algorithms.End-to-end training is done using a framework based on deep learning without using component-level annotation information,and fine-grained objects are detected using training models.By analyzing the characteristics of fine-grained images,mining the statistical characteristics of the target contour and improving the process of generating networks in the detection framework,an adaptive anchor mechanism is proposed and implemented to automatically fit the target contour closer to the target contour when solving different problems Anchor instead of manually set anchor point in the original frame.By using the pre-training model to improve the training effect,a more accurate fine-grained detection model is obtained.Finally,a fine-grained ship detection data set and system for internet data was implemented.In this paper,an application scenario for fine-grained warship detection problems is to collect images using crawler technology by capturing image data from the Internet.The SIFT feature is used to de-weight,filter and label the images.Contrasting the structure and characteristics of all kinds of mainstream datasets,we determine the principle of image annotation,the organization of document information,the construction of fine-grained data set of US warships,and carry out training and testing experiments.Taking the problem of ship detection as an example,this paper explores and realizes a complete method of fine-grained warship detection and identification under the method of deep learning.