Cosine Modified K-Means and Neural Network for Classification of Images

Abstract

A significant amount of data transfer happens daily through the internet in the form of images, PDFs, and videos. This exchange rate has increased tremendously during the COVID-19 pandemic. However, data transfer consumes a lot of internet bandwidth. It can be reduced significantly if there was a way to determine whether an image is compressed or non-compressed. While much research has been done on image compression in modern photography, detecting whether an input image is compressed or uncompressed has not been studied. This research aims to develop a new algorithm to classify images as compressed or uncompressed. The first step is to propose a new clustering method that uses two random centroids based on randomly selected pixels from the image. The Euclidean distance of each pixel from the randomly selected centroids is used to perform clustering. If clustering fails, then the cosine similarity method is used to perform clustering. This method is called the Cosine modified K-Means method. The SURF feature detection method is used to find the features of each image. Based on these extracted features, a neural network is trained. To test the algorithm, a random image is selected and passed through the network. The algorithm can classify the image as compressed or uncompressed. Precision, recall, classification accuracy, and error rate are calculated to evaluate the performance of the method.