Interactive tool for study and demonstration of lossy image transform coding

Abstract

Most images stored on a device are compressed by means of transform coding, whilst transform coding doesn’t actually compress the image it makes it possible for the quantization step to eliminate insignificant coefficients. This is because the transforms used for compression are designed to separate these insignificant coefficients from a given image. This software can be used as an educative tool to visualize the effects of lossy image transform coding in real time, before the image is actually compressed by encoding. The implemented transforms for this project include the DFT, DCT and DWT LeGall 5/3 Filter. The user is given the ability to manually interact with the transforms and quantize them using the provided scales. It services as an educational tool because the new size and PSNR is displayed immediately. Using this information, the user can compare transforms and derive his/her conclusion on the best performing transform. Besides comparing the transforms using the size comparison and distortion measure, intuitively one could conclude visually which is better by observing the blocking artifacts superimposed. This tool also provides freedom for experimentation, the user has the ability to use custom images with varying frequency. The DFT tool implemented can be used to filter both low magnitude or high magnitude coefficients, this allows one to study the importance each one holds. For the DCT besides quantization using the provided standard quantization matrix, custom quantization matrices can be inserted and saved. Again, using this option, one could set a quantization matrix to filter the high magnitude coefficients and study the effects. These features allowing certain freedoms on how an image is quantized lends itself to becoming an educative tool. The last tool implemented is the DWT, this tool provides the ability for sub-band decomposition and individual band quantization. A scale is provided that decomposes the image in sub-bands, this allows the image to be decomposed to the maximum level possible. Seven quantization scales have been included to allow individual band quantization. Upon quantization the new estimated size is displayed, the image is than reconstructed upon the push of a button. This allows the user to study the effects and importance of the bands, and which bands should be preserved for the least distortion and maximum compression.