컨벌루션 신경회로망과 ELM 분류기를 이용한 영상 분류

Abstract

This paper propose a method for combining of CNN(Convolutional Neural Network) with ELMC(Extreme Learning Machine Classifier) based on ReLU(Rectified Linear Unit) activation function. Basically, the proposed method consists of two phase development. First, CNN is employed to extract features of image database. Next, ELM is used to classify them. CNN is organized as a structure which repeats both convolution and pooling layer. First of all, the generated feature map will be downscaled to be its appropriate size through the pooling layer once a featured map is generated by convolution operation in the convolution layer. Consequently, the dimensionally reduced feature map are summarized in the fully-connected layer. The CNN used in this paper uses one of the architecture pre-trained model. The reason is that since hundreds of images have been already learned, it is easy to use for feature extraction and the feature values are classified as input data through the ELM classifier. And we use the activation function that is used to get the optimal output weight in the existing ELM classifier as the ReLU function used to compensate the disadvantage of the existing activation function in the CNN. The ELM is called SLFNs(Single hidden Layer Feed-forward Networks) because of its structure with one hidden layer, and it is a proposed method to complement some problems such as drop of learning rate caused by overfitting which is a disadvantage of back-propagation algorithm. Therefore, ELM show more exquisite performance capability at speed for classification than SVM(Support Vector Machine), Tree, Naive Bayes, and Discriminant Analysis, which are used in the existing machine learning classification methods. The ReLU function is a proposed method for solving the problem of gradient vanishing, which is a problem in existing neural networks. Moreover, it can be easily implemented by limiting negative values to 0. Also, it has the advantage that it allows the amount of computation to be reduced without the exponential operation used in the existing sigmoid or tangent function. When the ReLU function is used to obtain the output weights during the classification process of the ELM classifier, the performance is better than that of the conventional ELM classifier. These are three databases used in this paper: CIFAR-10, OxfordFlowers, Caltech-101. These are well-known databases and belong to the benchmark database for image classification. The CIFAR-10 database consists of 10 classes, 50000 training images and 10000 test images. The OxfordFlowers database is composed of 17 classes, 1360 images and 80 images for each class. The Caltech-101 database comprises 101 classes, and each of them contains different number of images. Because OxfordFlowers data and Caltech-101 data do not distinguish between training images and test images, 70% of the total images are used as training images and the rest are for the test images. As a result, we have been experimented in excellence of capacity of the proposed method, showing that it is improved approximately two to six percentages higher than the existing machine learning classification as well as better speed. It is also confirmed that the classification’s rates of the ReLU-based ELM classifier appears about two to five percentages higher than the existing activation function’s based ELM classifier.