基于深度集成学习的青梅品级智能反馈认知方法.

Fruit planting area and yield in China have reached the top level in the world. However, the lower processing level of the subsequent commercialization after fruit harvest is becoming one of the main factors to restrict the promotion of the added value and the international market competitiveness for the domestic fruit. Therefore, realizing the automatic classification of fruit grade has become an essential precondition of the modernization for the fruit industry in China. For the automatic classification method of fruit grade based on visible light technology, the working strength is considerably heavy and the cognitive effect is difficult to be satisfied, due to the susceptibility to the subjective factors for the man-made screening mode, such as experience. The corresponding machine screening mode based on the computer vision technology is susceptible to the drawbacks of the objective factors, such as traditional cognitive methods, and the classification result is also hence difficult to achieve satisfied effect. When the feature space and classification criteria are established once, they are un-updated, and are summarized as an open-loop fruit grade cognition mode for traditional machine judgment. Aiming at the defects, a greengage grade intelligent feedback cognitive method with cognitive result entropy measurement index constraint is explored, which imitates the human cognitive process with repeated comparison and inference from global to local. Firstly, under uncertain conditions and finite domain, from the information theory point of view, the greengage grade intelligent decision information system model is established by the representation of unstructured multi-level dynamic features with information completeness evaluation index. Secondly, the feature space data structure and classification criterion of greengage images with clear grade and feature mapping relationship are established based on adaptive structure-based convolutional neural networks and ensemble random vector functional-link net classifiers from global to local. Thirdly, based on the generalized error and generalized entropy theories, the entropy measurement evaluation index is established for the greengage image cognitive results. Finally, the intelligent operation mechanism of dynamic feedback cognition is established based on the measurement index constraint of uncertain process cognitive result. The average recognition accuracy of 1 008 greengage images for our proposed method is 98.15%. Such performance is 7.9% higher than the algorithm based on Gabor wavelet combined with principal component analysis and support vector machine. The performance of the algorithm based on color completed local binary pattern combined with the nearest neighbor classifier is also lower than that of the proposed method, and the average recognition accuracy of it is just 92.77%. Moreover, compared with the algorithm based on the wavelet descriptor combined with kernel principal component analysis and radial basis function neural network, the recognition accuracy of the proposed method is much better, although the running time is 0.7 s longer. The above mentioned conclusions indicate that the proposed method of adaptive structure convolutional neural networks and ensemble random vector functional-link net classifiers is suitable for the greengage grade machine screening recognition to replace the man-made screening mode, which can effectively enhance the generalization ability of the feature space and the robustness of the classifier. This study provides a reference for the rapid and accurate greengage grade machine cognition based on visible light. [ABSTRACT FROM AUTHOR]