Research On Image Recognition Algorithms Based On Primitive Feature Analysis

At present,artificial intelligence is undoubtedly one of the most popular technologies in the field of Internet.The development of AI benefits from the continuous promotion of machine learning,especially the breakthrough of deep learning.At present,the image recognition method based on deep learning neural network has achieved satisfactory results in recognition accuracy.However,with the increase of scene complexity of natural images and the decrease of euclidean distance difference between images of the same class,such methods often need to expand the training sample size,widen the network structure and deepen the network layer to achieve the ideal recognition effect.At the same time,because deep learning can use data to learn without understanding the data,such methods have poor stability in the face of undetectable disturbances in the data.The characteristics of the deep learning black box model make it impossible to explain the decisions made,so applications in many areas are limited by the lack of access to more reliable information.To address the above issues,this paper proposes an image recognition method based on primitive feature analysis.We design a semantic capsule fusion network(SCFN)for the recognition task,which includes a semantic Caps Net(SC)module and a simple convolutional neural network(CNN)module with auxiliary functions.In the SC module,we first detect the semantic primitives defined by human prior knowledge,analyse their features as semantic capsules,and then feed them into the Digit Caps layer of Caps Net.The CNN module is used to learn features which are difficult to describe by semantic primitives.By this way,a large number of simulation experiments prove that the proposed SCFN can achieve better image recognition performance with fewer training samples,higher training speed and stronger interpretability.The main work of this paper includes:(1)For the first time in this paper,the concept of image semantic primitives is proposed,and the perception method of image semantic primitives is introduced,which brings a new way for image coding and is more convenient for the completion of subsequent tasks such as image recognition.(2)Inspired by the capsule network,we proposed the concept of semantic capsule,and stored the semantic feature information of images in the semantic capsule.On the one hand,it was convenient for the overall construction of the subsequent identification network,on the other hand,it also made the network identification process have semantic characteristics,and the interpretability was strengthened.(3)The semantic capsule network is constructed,and after exploring,the semantic capsule fusion network is designed,which combines human prior knowledge with traditional neural network,and the fusion function is designed,so that the semantic capsule network and convolution neural network can work together.At the same time,they can exert their greatest advantages and make the network work better.(4)The whole simulation experiment is carried out on the MNIST handwritten character data set,taking into account the characteristics of the data set itself,which fully verifies the advantages of this paper's method in terms of training data and network parameters.