Enhanced extreme learning machines for image classification
Dongshun Cui · 2019
Image Classification is one of the critical computer vision tasks, and it is also the foundation of the related tasks like object detection/recognition/segmentation, which all need to identify the positive and negative samples.Methods proposed for image classification can be divided into two groups: traditional image processing based methods and modern machine learning based methods.The main difference between these two categories is how to extract the features, including the common features of the intraclass samples and the distinguishing features of the interclass samples.For traditional image processing based methods, researchers focus on extracting features manually after the mere observation of the samples.While for modern machine learning based methods, and researchers tend to focus on designing networks based on the training data, building a model, and evaluating the model on the test data.Among numerous machine learning methods, we choose the Extreme Learning Machines (ELMs) for our image classification applications.ELM has been proposed in more than a decade ago, as a single layer feedforward neural network, its primary objective is designing classification and regression models.ELM and its variants have been applied widely in the past years in many fields, including computer vision tasks, time sequence signals analysis, and so on.A basic ELM network consists of an input layer L 1 , a hidden layer L 2 , a output layer L 3 , and the connections between L 1 and L 2 , and L 2 and L 3 .From the view of data flow, it can be summarized as raw data or features of the samples are fed into the input layer, then flow into the hidden layer and the output layer by being operated with similar connections.The most prominent highlight of ELM is the weights and biases in the connections between L 1 and L 2 are randomly vi