Visual Saliency Detection Based on Multiscale Deep CNN Features.

Area of Science:

• Computer Vision
• Cognitive Science
• Artificial Intelligence

Background:

• Visual saliency is a key challenge in cognitive and computational sciences.
• Deep convolutional neural networks (CNNs) have shown great success in visual recognition tasks.

Purpose of the Study:

• To develop a high-quality visual saliency model using deep learning.
• To introduce a novel neural network architecture for extracting multiscale features for saliency detection.
• To create a new, large dataset for evaluating visual saliency models.

Main Methods:

• Utilized a neural network architecture with fully connected layers atop CNNs for multiscale feature extraction.
• Introduced 'deep contrast features' from the penultimate layer of the network.
• Integrated handcrafted low-level features with deep contrast features for enhanced robustness.
• Constructed a new dataset (HKU-IS) of 4447 challenging images with pixelwise saliency annotations.

Main Results:

• The proposed method achieved state-of-the-art performance on public benchmarks.
• Improved F-measure by 6.12% on DUT-OMRON and 10% on HKU-IS datasets.
• Reduced mean absolute error by 9% on DUT-OMRON and 35.3% on HKU-IS datasets.

Conclusions:

• Deep convolutional neural networks can effectively learn high-quality visual saliency models from multiscale features.
• The proposed 'deep contrast feature' is a discriminative high-level feature for saliency detection.
• The new dataset and proposed method advance research and evaluation in visual saliency.