Research On Fine-grained Bird Recognition Based On Convolution Neural Network

Artificial intelligence is gradually coming into the eyes of people,which indicates that the pace of intelligent human life is gradually accelerating.Among them,as an important part of artificial intelligence,deep learning and image recognition have drawn more and more attention.The general classification of image recognition can no longer meet the human requirements for the retrieval accuracy.Fine-grained identification is gradually becoming the most active research area.The phenomenon of "semantic gap" between local features and high-level semantic features makes the general recognition classification methods can not effectively learn the local key features and correctly grasp the connection between them,so that the recognition result is not satisfactory.In this thesis,the problem of bird's fine-grained recognition based on convolutional neural network is studied and analyzed.The research work includes the following aspects:(1)First this paper introduces the research background of fine-grained recognition,and then explains the importance of fine-grained recognition research,and analyzes the fine-grained recognition problem,mainly summarizes and summarizes the research status at home and abroad.Then we discuss the research status of the convolutional neural network.And the paper proposes the recognition algorithm,which is based on convolutional neural network.Finally we describe the convolutional neural network structure,characteristics and algorithms,it provides the theoretical basis for constructing the fine-grained network model identification of birds.(2)In order to obtain the local area with rich details,this paper proposes a fine-grained identification method based on semantic detection of Faster RCNN,constructs an end-to-end network,and uses K-NN top-down geometric restriction method to generate seven regional candidates Region,feature extraction is performed on the candidate region according to the constraints,this method greatly reduces the number of candidate regions on the order of magnitude.Experimental results show that this method is suitable for bird fine-grained identification and can improve the recognition accuracy.(3)With the increase of the network convolution depth,the RoI pooling layer often can not achieve satisfactory results,leading to the loss of important features.In order to solve this problem,this paper proposes a Faster RCNN model based on multiscale feature connections.The method is to connect the features from multiple lower-level convolution layers to a pooling layer respectively and to merge the learned features and Normalized.The experimental results show that the scheme of multiscale feature connection in training can better learn global and part features and help to improve the recognition accuracy.