Compression through extraction of learned parameters from images in de-correlated image space

Image compression is a class of algorithms that reduces the storage space requirement for a digital image. Lossy image compression techniques achieve higher compression but the visual quality of the decompressed image is degraded many times. Decompressed images lose their visual appeal due to compression artifacts. These compression artifacts are introduced due to the quantization step of the compression phase. We developed a lossy image compression technique that works on the spatial domain and de-correlated color model. For the luminance channel compression, the modified Vector Quantization method is used. In the case of chrominance channels, a feature vector is built for each pixel using the neighborhood statistics and cluster information of the pixel. For all the pixels of the image, using these feature vectors, a training dataset is formed. For the training of an artificial neural network (ANN), a feature vector of a pixel is used as the input and its respective chrominance value is used as the target output. Two training datasets are used to train two ANNs separately—one for the Cb channel and one for the Cr channel. These two trained ANNs are stored as the compressed form for the chrominance channels. During the decompression process, first, the luminance channel is reconstructed. Later, for each chrominance channel, the respective trained ANN predicts the chrominance values for each pixel. Thus, the whole image is reconstructed. The method has been tested on the benchmark images and also color images from the UCID v.2 database. The experimental result shows that the method successfully avoids the blocking artifacts in the reconstructed images.