SSTop3: Sole-Top-Three and Sum-Top-Three Class Prediction Ensemble Method Using Deep Learning Classification Models

Abstract

Computer Vision (CV) has been employed in several different industries, with remarkable success in image classification applications, such as medicine, production quality control, transportation systems, etc. CV models rely on excessive images to train prospective models. Usually, the process of acquiring images is expensive and time-consuming. In this study, we propose a method that consists of multiple steps to increase image classification accuracy with a small amount of data. In the initial step, we set up multiple datasets from an existing dataset. Because an image carries pixel values between 0 and 255, we divided the images into pixel intervals depending on dataset type. If the dataset is grayscale, the pixel interval is divided into two parts, whereas it is divided into five intervals when the dataset consists of RGB images. In the next step, we trained the model using the original dataset and each created datasets separately. In the training process, each image illustrates a non-identical prediction space where we propose a top-three prediction probability ensemble method. Top-three predictions of newly generated images are ensemble to the corresponding probabilities of the original image. Results demonstrate that learning patterns from each pixel interval and ensemble the top three prediction vastly improves the performance and accuracy and the method can be applied to any model.