Digital Color Image Compression : With Real and Complex Artificial Neural Networks

Abstract

Neural networks are an exciting and evolving branch of machine learning, but they are not limited to just artificial intelligence. Recently, they have been used to compress and even add digital watermarks, or copyright signatures, to digital images. Typically, neural networks use real numbers for their computations, but researchers have also experimented with using complex numbers and quaternions as the basis of these networks. Our research investigates the use of quaternion-valued neural networks implemented in Java for the purposes of digital image compression and watermarking. The benefits of using quaternion-valued over real-valued neural networks include faster network training time, better color compression/recovery, and less processing power/memory required for the computation.