Deep Attentive Skip Connection Model and Gaussian Priors based Visual Saliency Prediction

Predicting human visual attention mechanisms using deep learning has recently attracted a great amount of research interest in the computer field. The visual system of humans serves as a filter to guide greater attention to the most attractive and interesting objects for subsequent processing. With the recent widespread use of deep learning algorithms, the problem of predicting saliency has seen a tremendous improvement. In this study, a saliency prediction scheme is proposed that includes an Attentive Convolutional Long Short-Term Memory network (Attentive ConvLSTM) that repeatedly refines saliency features by focusing on key spatial locations. The input picture is first utilised to extract a series of feature maps using VGG16, a dilated convolutional network. After that, to improve predictions Attentive ConvLSTM architecture is proposed in the form of Skip Connection with VGG16 feature map which successively focuses on different spatial locations of a stack of features. Then the predictions are merged with various learned priors to illustrate the tendency of humans to fix the centre region of the image. The success of this proposed Deep Attentive Skip Connection with Gaussian Priors Model (DASC-GP-Model) is evaluated on MIT saliency benchmark dataset and this model achieves 90.97% of accuracy. From the results, it is observed that this proposed network can predict strong saliency values on individuals, things, faces, and other important cues and also it creates virtuous saliency maps.