AE and SAE Based Aircraft Image Denoising

Abstract

Images are corrupted during transmission and acquisition. De-noising is an important image restoration operation which determines the accuracy of interpretation and recognition stages. Time and often traditional methods have been used for image de-noising. Lately, there has been considerably interest on learning aided image de-nosing. As deep learning has lately been established as the most efficient learning aided mechanism, it is increasingly being used for a range of image processing and computer vision applications. This paper focuses on the design of Auto-encoder (AE) and Stacked Auto-encoder (SAE) based approaches for de-noising of certain military aircrafts as part of an automatic target recognition (ASR) system. Five image types are taken for the work which are mixed with Gaussian, Poisson, Speckle, Salt and Pepper noise. For each of these image sets signal to noise ratio (SNR) variation between -3 to 10 dB are taken. Experimental results have show that the SAE based approach is more reliable despite showing higher computational latency.