Ensemble Learning For Material Recognition And Segmentation With Convolutional Neural Networks

With the rapid development of artificial intelligence,the concept of intelligent life is becoming more and more popular.And as the basic researches in the field of computer vision perception research topic,material recognition and segmentation have drawn more and more attention from researchers.The material recognition and segmentation technology are widely applied to the automatic classification of household waste and the indoor intelligent navigation,at the same time,the material recognition and segmentation are also the preprocessing procedures in many computer vision tasks.Therefore,it is an important research topic for improving the accuracy of material recognition and improving the result of material segmentation.However,there are varieties of materials in our daily life,it is a challenging task to choose the distinguish feature and improve the accuracy of material recognition,which is also a hot topic for researchers.With the continuous researches on the material recognition and the rapid development of deep learning,there is a breakthrough on the material recognition with the convolutional neural network and the large-scale database.This paper proposed an algorithm of material recognition that based on convolutional neural network and ensemble learning method.And we apply the algorithm to material segmentation with conditional random fields.The main research work of this paper is as follows:(1)Material recognition based on convolutional neural network and ensemble learning.Firstly,we train the CNN model to extract features,and the knowledge-based classifiers are trained on these features.And then we proposed the ensemble learning methods on the probability-level.We combine these knowledge-based classifiers with our methods,and try to improve the recognition accuracy by increasing the probability of correct category and reducing these probabilities of error categories.Meanwhile,a weight learning method is proposed to set weights for each of knowledge-based classifiers.(2)Material segmentation based on modified CNN and the conditional random fields.In this part,we take the modified CNN as the sliding window to identify the material category for each pixel of the test image.We will get the material category probability map,and then combine the conditional random fields to achieve the purpose of material segmentation.(3)Design and realization the material recognition and segmentation system.We apply above methods to material recognition and segmentation into the system.The system can recognize or segment the material category for the imported image with the trained models.The results will be showed in the specified locations.All of the research works have achieved the desired purpose.Experiments carried on the public database show that the recognition accuracy of the material recognition based on the convolutional neural network and ensemble learning is improved,and simultaneously the result of the segmentation is also getting enhanced.