Research On Image Recognition Based On The Fusion Of Semantic And CapsNet

Image recognition uses computer to process and analyze image information to divide the target categories contained in the image.The application of image recognition in life is quite common,such as fingerprint recognition,face recognition and traffic sign recognition etc.Convolutional neural network recognition method has excellent performance in image recognition because of its powerful feature extraction ability,but the methods are completely data-driven,resulting in a large training sample demand,long training time,difficulty in adjusting the parameters,etc.This thesis proposes an image recognition method which fuse semantic and CapsNet.Image semantics are the structured knowledge description by visually receiving image information,which is comprehensible and interpretable.By fusing the semantics of the image and CapsNet,the utilization of training data is improved.The proposed method achieves better recognition results with less training set.The proposed method can effectively alleviate the problem that the deep learning model relies on a large number of training samples.The main works of this paper are as follows:(1)This thesis proposes a recognition network based on semantic,it simulates the human's targets cognition: People recognize different categories of objectives by the semantic.Semantics are different categories of objectives' specific concept descriptions.Different semantics have a certain hierarchical relationship.Semantics are often composed of several sub-semantics.Sub-semantics can be divided until they cannot be disassembled which called semantic primitives.This thesis constructs a semantic network to describe the relationship between semantic primitives and abstract semantics.Through the recognition of semantic primitives and the combination of different semantic primitives in specific abstract semantics,specific targets are identified.However,the semantic description feature is based on human knowledge which has limited ability to portray details,and the image collected in the natural environment have various degrees of interference such as occlusion and insufficient illumination,thus the extracted semantic description feature isn't sufficient to recognition the target.By designing the loss function,fusing semantic network and capsule network,which can enhance the ability of detail description and improve the training set utilization.The proposed method achieves a better recognition effect in the case of training using small sample training set.(2)Analyze the dynamic routing algorithm in capsule network proposed by Hinton,three improved algorithms are proposed.The experiments of the proposed three different algorithms are performed on MNIST respectively.The experimental results shows the proposed three algorithms have achieved better recognition accuracy than the three iterations of the dynamic routing algorithm.The proposed adaptive clustering algorithm is tested on CIFAR10 and GTSRB.The result shows that the proposed adaptive clustering algorithm solves the problem that the dynamic routing algorithm has a large amount of parameters,high computational complexity and the optimal number of iterations is difficult to determine because of the iterative processes.The proposed adaptive clustering algorithm has the advantages of less parameters,wider applicability and scalability.