Lossy Compression of Noisy Images Using Autoencoders for Computer Vision Applications

Computer vision gets a lot of concern in everyday applications across several domains such as industry, surveillance, medicine, military, and autonomous vehicles. In this context, deep neural networks (DNNs) have extraordinary capabilities and are widely used. Convolutional neural networks (CNNs), for the most part, comprise a prevalent class of DNNs analyzing visual imagery. However, CNN’s performances depend entirely 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 inhibit a proper CNN classification of images. The second issue is the throughput available to transmit the large amount of data between capture sensors and units processing CNNs. A seamless broadcast can be ensured by compression techniques that reduce data size while affording computer vision algorithms’ required quality. The outcomes of lossy compression of the ground truth (GT), which is clear/noiseless/undistorted, and noisy images differ. Hence, this work is first devoted to studying noisy image compression using particular autoencoders. The authors typically raise the question of CNNs’ resilience to such compression. Second, the super-resolution-based learning architecture relying on CNNs improves error robustness and can achieve better classification performances.