Investigating U-Net's Reliability in Saliency Detection with Salt and Pepper Noise

In this digital age, images and videos are the main source of conveying information. People can concentrate very well on the most crucial areas of an image or video. Emulating the human visual system to determine the most noticeable object is a popular area of study in the field of computer vision and image processing. Salient objects are those that stand out the most in the image. Numerous applications, including automated cropping, video summarizing, compression, and others, are greatly impacted by salient object detection approaches. Recent rapid advancements in artificial intelligence have led to the proposal of multiple deep learning-based saliency detection models by researchers. Nevertheless, the majority of approaches concentrate on saliency detection in noise-free environments. Since noise can affect images or videos, it can be difficult to handle the noise and identify the important object in noisy conditions. The detection of the presence of noise in the image or frame is also a challenging task. Here in this work, the author(s) tried to analyze the effect of salt and pepper noise in the saliency detection process. In the proposed work first, the author(s) developed a noise classifier that identifies the presence of salt and pepper noise. Further, the existing noise reduction algorithms are used for denoising purposes, and finally for saliency detection lightweight U-Net model is utilized to identify the most prominent object. In a clean, noisy, and denoised environment, the lightweight U-Net model's performance is evaluated in this work.