Concept Learning Method Of Image Understanding

In the past few years,many image understanding methods based on deep learning have developed rapidly,especially in the field of character recognition.However,these methods are faced with many problems,such as large amount of training data,many parameters,large consumption of computing resources,and the generalization ability is also weak.Concept learning is a kind of learning method like human beings.human beings can learn efficiently under the condition of few samples and prior knowledge,so this has strong generalization ability.Different from the usual deep learning model,the concept learning method only needs one training sample to build the concept model on the basis of a certain prior knowledge.In view of the shortcomings of deep learning method,we take Chinese handwritten characters and industrial character images as experimental objects.On the basis of reasonable character segmentation and character stroke representation,we use concept learning method to build character concept model,and further realize character image understanding.The main work of this paper is as follows:1.Handwritten Chinese characters may be written in a continuous manner,and character images in the industrial environment may be printed on the surfaces of various materials with complex and uneven shapes,resulting in incomplete and contact characters.According to the features of contact characters,the fuzzy segmentation strategy of Support Vector Machine(SVM)and Particle Swarm Optimization(PSO)are used to segment contact characters reasonably.2.Because the concept learning method needs to split the whole into different parts,it needs to split and represent the character image reasonably.In the process of character stroke segmentation and extraction,redundant feature points are easy to produce,which leads to the failure of stroke segmentation.In this paper,the method of removing redundant feature points based on stroke width and the method of correcting distorted skeleton based on distance histogram are proposed,so as to achieve the reasonable extraction of character stroke.In addition,because concept learning is based on statistical learning method,strokes need to be represented by different parameters,and the probability distribution that parameters obey is calculated.3.According to the representation of strokes,the order of strokes and the connection relationship between strokes,a single character conceptual model is established.On the basis of the established conceptual model,the Monte Carlo Markov chain sampling method is used to generate the fitting model by changing the relevant parameters of the samples.The probability value of fitting model is calculated to determine the category of target samples.At last,we can use naive Bayes decision to distinguish extremely similar characters.4.In order to verify the applicability of the method proposed in this paper,we not only evaluate the industrial image database collected in the industrial scene,but also carry out experiments on the contact,incomplete and similar character data sets and obtain good recognition accuracy.Good performance is also on public data set ICDAR 2013 and NIST SD 19.