Image Compression Using Generic Vector Quantizer Designed Using Transform Coding: The Quality Analysis Perspective

This paper presents a novel idea of designing codebook named as generic codebook and generic codebook with error for compression of gray images. The codevectors are generated by evaluating the characteristics of the specific image sub samples, which are determined through rigorous mathematical operations and training the selected samples by Kohonen´s self organizing feature map (SOFM) artificial neural network with adjustable learning rate and initializations conditions. The testing is done with variety of images, especially high frequency images. The codebooks are also designed by applying discrete cosine transform (DCT) and discrete wavelet transform (DWT) to the code vectors of generic codebook. The images are compressed using DCT and DWT based codebooks. The compression performance is evaluated by using objective and subjective quality measures. Peak signal to noise ratio (PSNR) is widely used measure for determining the quality, which is deceptive many times, other quality measures such as image fidelity, mean structural similarity index (HVS based) spatial frequency measure (SFM) and spectral activity measure (SAM) which deal with spectral activity and coding complexity are also used along with PSNR. The quality is also determined using no reference quality measure called score. Finally, the performance of three vector quantizers with error are compared based on quality analysis, coding complexity and hardware implementation.