Visual Saliency Prediction Using Deep Learning

Abstract

Visual saliency prediction is one of the most trending technology that may also leads to the future. The object detection is one of the main concepts behind this saliency prediction. So high level visual features are augmented with contextual information such as text, images etc. One method of prediction is by using the convolution neural network with the encoder-decoder architecture. This method achieves accurate, efficient and competitive result. The visual saliency prediction is used as a highlighter for the meaningful regions in the scenes. The other method is done using the cues which can reduce the weight of the network and this network takes 224*224 images. The other identification is done using Human Visual System with the usage of the architecture of fully connected convolution network. The other way of prediction is through the Deep Neural Network which has more trainable data and so some difficulties faced through this process can be removed. Another approach is done using neuroscience principles which includes a data driven approach by Convnet. Euclidian distance is measured for the performance of the network. All these methods are done using the publicly available datasets and some of them are MIT1003, MIT300, CAT2000, DUTOMRON etc.