Signal Extraction for Classification of Noisy Images Compressed using Autoencoders
Dorsaf Sebai, Nour Missaoui, Asma Zouaghi · Computer Science Research Notes · 2021
The world is experiencing an increasing boom in computer vision.This is more and more used in many domains such as robotics, medicine, industry, security systems, etc.In this context, Deep Neural Networks (DNNs) have great capabilities and are widely used.Convolutional Neural Networks (CNNs) present a particular class of DNNs that is most commonly leveraged to analyzing visual imagery.However, CNN performances completely depend on two main issues.The first issue is related to the quality of the images generated by capture cameras.All images captured by remote sensors and modern imaging systems are practically noisy, which can prevent the image from being correctly classified and identified by a CNN.The second issue is the throughput available for the transmission of the large amount of data between capture sensors and units processing CNNs.A seamless transmission can be ensured by compression techniques that help reducing the size of data, while affording the required quality for computer vision algorithms.Since lossy compression of noise-free and noisy images differ from each other, this work firstly raises the question of CNNs resilience to noisy images compression using the particular autoencoders.We secondly propose a method that aims to improve this resilience so that CNNs can achieve better classification performances.The compressed noisy images are passed, as a test set, along a model that is learnt from a noise dataset.The subtraction of the so captured noise from the noisy images is then performed to extract the useful signal to classify.This will be first work, where we learn the autoencoder from the noise sample, and not the noisy sample, while denoising.Obtained results prove the efficiency of the proposed method.