Research On Similar Greengage Grade Intelligent Cognitive Method Based On Semi-supervised Learning

Chinese fruit planting area and yield are in the forefront of the world,but the level of commercialized processing and processing ability after fruit picking are the main factors that restrict the sales of the international market.As a kind of medicine and food resources with multiple health functions,greenage is well received by the masses.The performance of manual sorting mode is easily affected by subjective factors such as operator's experience and sense of responsibility.Therefore,automatic sorting based on machine vision is one of the main technical means of fruit grading.However,the large number of similar samples and the difficulty in obtaining sample labels are the main reasons for the poor performance of machine sorting.So,how to use a large number of unlabeled samples and obtain local subtle differences are the key to distinguish similar samples.Aiming at the inherent defects of supervised learning mechanism and deep neural network structure,this dissertation proposes a deep intelligent cognitive method based on semi-supervised learning with uncertain cognitive result entropy measurement index constraints,which is applied in similar greenage classification.The main work of this dissertation is list as follows:(1)In orde to solve the difficulty in obtaining similar green plum image samples with labels,the graph-based semi-supervised learning method uses less labeled data to label a large number of unlabeled data,and enlarges the training sample set of greenage image with high reliability sample,which reduces the labor intensity.(2)In orde to solve the problem of insufficient characterization of differentiated features between similar greenage images,a deep adaptive neural network is constructed under the constraints of maximum information entropy performance index,and the multidimensional fully differentiated feature space of similar green plum images from global to local is obtained.Based on the separability metric measure index and variable precision rough set,under the condition of limited domain uncertainty,the cognitive intelligent decision information system model is established from the perspective of information theory.the simplied multi-level fully differentiates mapping relationship with the similar greenage grade is extracted.The multi-level fully differentiates feature space data structure which has clear reflection with similar greenage grades clear is structed.(3)In orde to solve the shortcomings of generalization ability of softmax layer in deep neural network.An integrated RVFL classifier is used to construct the classification cognition criterion for similar greenage images with reduced multi-level fully differentiated feature space.Based on the forward propagation and reverse update according to alternating optimization strategy.Alternate optimization strategy,global iterative to update the parameters of deep neural network and integrated RVFL classifier.(4)Aiming at the problem of posterior statistical evaluation of uncertain cognitive results and the poor applicability of fixed feature space.a measurement index of error entropy based on generalized error and generalized entropy for uncertain cognitive results of similar greenage grade.The reliability of uncertain cognitive results of similar green plum grade is evaluated in real time.The cognitive accuracy,network level and neutralization factor are adjusted alternately by self-optimization to update the reliability of the feature space,classification cognitive criteria,calibration of unlabeled sample and then re-recognize the similar green plum image samples with low credibility.In order to verify the effectiveness of the proposed algorithm,3000 similar greenage image were simulated,and the average cognitive accuracy rate reached 98.26%,which provids theoretical support for the subsequent design of online sorting system.