University essay from Umeå universitet/Institutionen för datavetenskap

Author: Alexander Hjältén; [2019]

Keywords: ;

Abstract: th‘e quality of an image plays a role in how well it can be correctly classi€fied by an image classifying neural network. Artifacts such as blur and noise reduces classi€filability. At the same time it is oft‰en motivated to reduce fi€le sizes of images which also tends to introduce artifacts and reduce their quality.‘ This leads to a trade-off‚ between having small fi€le sizes and achieving high classi€fication accuracy. ‘The two main approaches of reducing €file sizes of images is to either reduce the number of pixels in them via image scaling or to use less data to represent each pixel via compression.Th‘e e‚ffects of these two approaches on image classi€fication accuracy have been studied independently.In this study the eff‚ects of combining image scaling and compression in regards to image classi€fiability is examined for the €first time. Images are downscaled using fi€ve popular methods before being compressed with di‚fferent magnitudes of JPEG compression. Results are evaluated based on the fraction of the treated images that are correctly classifi€ed by the classi€fier as well as on the image fi€le sizes.Th‘e results shows that the scaling method used has signifi€cant but weak e‚ffect on image classifi€ability. ‘Thus the choice of scaling method does not seem to be critical in this context. ‘There are however trends suggesting that the Lanczos scaling method created the most classi€fiable images and that the Gaussian method created the images with highest classi€ability to fi€le size ratio. Both scaling magnitude and compression magnitude were found to be be‹tter predictors of image classi€fiability than scaling method.

  AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)