PROJECT TITLE :

Deep Visual Saliency on Stereoscopic Images

ABSTRACT:

Quality of stereoscopic 3D images has been demonstrated to have a significant impact on visual saliency in S3D images. As a result, this dependency is critical in predicting image quality, restoring visuals, and reducing pain, although it is still challenging to forecast such a nonlinear relation in images. Furthermore, algorithms that are designed to detect visual saliency in perfect images may fail miserably when confronted with warped ones. A deep learning approach called Deep Visual Saliency (DeepVS) is examined in this study to improve the accuracy of a saliency prediction even when distortion is present. We propose seven low-level features (contrast, brightness, and depth information) extracted from S3D image pairings and use them in the context of deep learning to detect visual attention adaptively to human perception, as shown by psychophysical research. To extract distortion and saliency information, it appears that low-level features play a role. To build saliency predictors, we use a regression and a fully convolutional neural network to simulate the human visual saliency. The findings of extensive trials show that the projected saliency maps are up to 70% linked with human gaze patterns, emphasising the importance of hand-crafted features in S3D saliency detection.