Salient object detection from low contrast images based on local contrast enhancing and non-local feature learning

摘要

Salient object detection can facilitate numerous applications. Traditional salient object detection models mainly utilize low-level hand-crafted features or high-level deep features. However, they may face great challenges in the nighttime scene, due to the difficulties in extracting well-defined features to represent saliency information from low contrast images. In this paper, we present a salient object detection model based on local contrast enhancing and non-local feature learning. This model extracts non-local feature combines with local features under a unified deep learning framework. Besides, a deeply enhanced network is employed as a preprocessing of the low contrast images to assist our saliency detection model. The key idea of this paper is firstly hierarchically introducing a non-local module with local contrast-processing blocks, to provide a detailed and robust representation of saliency information. Then, an encoder-decoder image-enhanced network with full convolution layers is introduced to process the low contrast images for higher contrast and completer structure. As a minor contribution, this paper contributes a new dataset, including 676 low contrast images for testing our model. Extensive experiments have been conducted in the proposed low contrast image dataset to evaluate the performance of our method. Experimental results indicate that the proposed method yields competitive performance compared to existing state-of-the-art models.