Abstract
Cosaliency detection aims to identify common salient objects within a group of related images. This paper introduces a novel framework that utilizes Stacked Denoising Autoencoders (SDAEs) to address the challenges of varying object appearances and limited training data. The approach consists of two main components: transferring intrasaliency priors from large auxiliary datasets and mining deep intersaliency patterns via unsupervised reconstruction residuals. Experimental results on standard benchmarks demonstrate that this method significantly outperforms traditional shallow-feature-based approaches.
Introduction
In computer vision, saliency detection identifies regions that naturally attract human attention. Cosaliency detection extends this to image groups, requiring the identification of objects that are both salient within their respective images (intrasaliency) and common across the entire set (intersaliency). Traditional methods often struggle when common objects exhibit significant variations in pose, lighting, or background. Zhang et al. (2016) proposed a deep-learning-based solution to capture these complex semantic commonalities.
Methodology
The proposed framework follows a three-stage pipeline to generate final cosaliency maps, integrating both intra-image and inter-image cues.
1. Intrasaliency Prior Transfer
To avoid the limitations of hand-crafted rules, the authors employ transfer learning. They train an SDAE on a large auxiliary dataset (ASD) to learn abstract "contrast" priors—measuring how distinct a region is from image boundaries. This "transfer machine" allows the model to predict saliency in new image groups without needing group-specific labels.
2. Deep Intersaliency Mining
The core innovation is the use of SDAEs for unsupervised pattern mining. The model takes salient regions from all images in a group to self-train a new SDAE. Because the network learns to represent the group's commonalities, regions that fit these patterns yield low reconstruction residuals. These residuals serve as a robust metric for "Deep Intersaliency," capturing semantic similarities that shallow features miss.
3. Shallow Intersaliency and Fusion
The deep mining results are integrated with traditional "shallow" features (like low-level color and texture matching) and the transferred intrasaliency priors. A weighted linear combination of these cues produces the final, refined cosaliency map.
Experimental Results
The model was evaluated on the Image Pair and iCoseg datasets. On the challenging iCoseg dataset, the method achieved an Average Precision (AP) of 0.878 and an Area Under the Curve (AUC) of 0.969, surpassing contemporary state-of-the-art models.
Qualitatively, the framework demonstrates superior robustness in suppressing cluttered backgrounds—such as grass or spectators in sports photos—that might appear frequently but are not the primary objects of interest.
Conclusion
By combining transfer learning for intrasaliency and unsupervised deep mining for intersaliency, this framework provides a powerful tool for identifying common ground in image sets. While the per-group training of SDAEs introduces some computational overhead, the ability to capture deep semantic features marks a significant advancement over previous shallow methods. Future work may focus on more sophisticated integration techniques and extending the framework to video sequences.
Summary updated