Adaptive digital image data compression using RIDPCM and a neural network for subimage classification

Abstract

Recursive Interpolated Differential Pulse Code Modulation (RIDPCM) is a fast and efficient method of digital image data compression. It is a simple algorithm which produces a high quality reconstructed image at a low bit rate. However, RIDPCM compresses the entire image the same regardless of image detail. This paper introduces a variation on RIDPCM which adapts the bit rate according to the detail of the image. Adaptive RIDPCM (ARIDPCM) is accomplished by dividing the original image into smaller subimages and extracting features from them. These subimage features are passed through a trained neural network classifier. The output of the network is a class label which denotes the estimated subimage activity level or subimage type. Each class is assigned a specific bit rate and the subimage information is quantized accordingly. ARIDPCM produces a reconstructed image of higher quality than RIDPCM with the benefit of a further reduced bit rate.