wganBCS : Block-wise image compressive sensing and reconstruction model using WGAN-gp network to eliminate block effects

In the past decade, compressive sensing has been successfully applied to many image processing tasks based on an empirical observation that most image signals are sparse in a certain domain. Several image compressive sensing methods have been proposed these years including the block compressive sensing (BCS) which uses a relatively smaller sensing matrix to raster-scan the entire image to acquire the signal block by block, which can greatly reduce the memory consumption. However, BCS still suffers from two issues. One is that block-wised sensing causes heavy block effect on the reconstructed image, which leads to degradation in the image quality metrics. Besides, the sensing matrix usually needs to be handcrafted, which requires a huge amount of work. In this paper, we propose a WGAN-gp based image compressive sensing and reconstruction model using generative adversarial networks which have achieved a great success these years on dealing with computer vision tasks, e.g. super-resolution and deblurring. More precisely, we train a plain convolutional neural networks to perform the block-wised sampling and initial reconstruction. Then we use a generative adversarial network which provides fine reconstruction on the initial reconstructed images to eliminate the block effect to improve the final reconstruction quality. Experimental result shows that our model is superior both in visual quality and in the image quality metrics compared to the traditional image compressive sensing methods.