Convolutional autoencoder for saliency dectection

Abstract

Salient regions are known to be important for the human visual system. In computer vision, the associated saliency detection task has received an increasing interest in the recent years. Despite the substantial progresses made following the introduction of deep learning approaches, many challenging issues remain. Indeed these methods often process local regions within images separately and as a result have difficulties capturing global patterns. In addition saliency maps are often blurry around the boundary of the salient object. To address these issues, we propose a convolutional autoencoder which first extracts the information contained in input images using a convolutional network (encoder) before generating output saliency maps using a deconvolution network (decoder). Our network operates on whole image directly so that global patterns are captured with ease. To reduce blurriness in the output we add direct connections between the encoder and the decoder. Finally the inclusion of a novel contrast penalty term helps further improve the sharpness of the output, especially around the edges of salient objects. We compare our method with six other state-of-the-art algorithms on three widely used benchmarks, where it shows equivalent performance. We also conduct four experiments to better understand how encoding and decoding are done.