Obtaining paired noisy-clean images for various types of corruption is challenging; however, a noisy image can be viewed as the superposition of two distinct signals. Drawing inspiration from this concept, we address the problem of image purification by focusing on separating these signals to recover accurate classifier decisions. We introduce a dual-tail convolutional autoencoder designed to isolate the noise signal from the clean image. This architecture is engineered to simultaneously generate the additive noise pattern and the original clean signal. We conducted extensive experiments across various types of natural image noise with differing severity levels under both seen and unseen conditions. The results demonstrate that the proposed unique architecture effectively manages multiple noise types and significantly improves object recognition performance, which is severely impacted by image corruption. For example, Salt & Pepper noise reduces ResNet’s accuracy on CIFAR10 from 91.81% to 20.48%, however, the dual-tail signal separator restores it to 91.61%. Additionally, the proposed method outperforms state-of-the-art approaches, uncovers connections between different corruptions, and, being cost-effective, has the potential to enable safe and secure AI deployment on low-cost devices.