A Study on the Generalized Normalization Transformation Activation Function in Deep Learning Based Image Compression.

Imagecompressionhasalonghistory, beingwidelyappliedinnumerousapplications. Yet, the commonly applied approach is based on traditional signal processing, such as JPEG. The lack of learning process in the approach limits the performance of those methods. Deep learning, recently, shows great performance in image compression, whose success may be attributed to various factors. Activation function is known as one of these influential factors. The present work is devoted to analyze the different effects of various activation functions, and the experimental results suggest that the generalized divisive normalization (GDN) is probably the best activation function in deep image approach-based image compression.