A feedforward neural network compression with near to lossless image quality and lossy compression ratio

In this paper, a novel image compression algorithm is proposed by introducing the feedforward neural network (FFN) instead of the quantization block of a general image compression algorithm. In this new method, after the spatial information of the target image is transformed into the equivalent frequency domain, the FFN stores each of the transformed coefficients in the network synaptic weights. By storing just the network weight values, the amount of information need to be retained for decompression purpose is much lesser compared to lossless method which stores information pertinent to each pixel in the image. As a consequence, a better compression ratio can be achieved by the FFN compression method as compare to lossless compression. Furthermore, during the decompression stage the FFN is capable of reproducing every single frequency component (coefficient values) with small margin of error due to the fact that no information is reduced unlike in lossy methods where some psychovisual redundancies are removed in the quantization. Results show that this new proposed compression algorithm (FFN compression) is capable of achieving the competitive advantage of lossy methods which is the compression ratio without compromising the image quality, the advantage of lossless methods.