Image quality in image classification: Design and construction of an image quality database

Abstract

Image quality affects automated classification of images from process camera monitors. The objective of this work was to obtain a database of reference images that could enable automated, customized image quality modification to improve classification of new images. Here, images from an extruder monitor were to be classified as either showing or not showing contaminant particles in a polymer melt. A novel task-based definition of image quality was important to this work: image quality was defined in terms of the probabilities of a particle being present and not being present in the image as assigned by a Bayesian classification model. Image quality was optimized using a Nelder Mead Simplex search. The optimized image was classified using another Bayesian model to demonstrate the improved classification performance. The resulting reference image database consisted of image similarity attributes describing each raw image and the corresponding quality improvement instructions from the optimization. The next step is to use the database to improve the quality and classification of new images.