A Convolutional Encoder-Decoder Network With Skip Connections for Saliency Prediction

In this paper, we propose a novel convolutional encoder-decoder network with skip connections, named CEDNS, to improve the performance of saliency prediction. The encoder network utilizes the DenseNet model as the stem network to extract abundant hierarchical features from input images. Subsequently, a decoder network is designed to sufficiently fuse the hierarchical features to predict saliency more accurately. Between the encoder and decoder, skip connections are employed to transfer hierarchical features produced by the former to the latter. Furthermore, the model can be trained in an end-to-end manner which is beneficial for both training and inference. The experimental results on various benchmark datasets, SALICON, MIT300, and CAT2000, show that the proposed model achieves state-of-the-art performance on several key metrics.