Exemplar Based Image Salient And Co-salient Object Detection

Saliency detection and co-saliency detection are important issues in computer vision.Heuristic saliency methods usually fail to detect proper salient object in complex images,and usually ignore the utilization of labeled dataset.Generative saliency methods via deep neural network depend on large amount of training data,while require expensive experiment and plenty of training time.In order to conquer these problems,we adapt an exemplar based method for saliency and co-saliency detection.In this paper,we exploit labeled data by retrieving labeled images from retrieval set with similar foreground to the query image.Utilizing these foreground of retrieval images and approximate background of query image,we can attain meaningful exemplars with proper feature extractor.We use these exemplars to train a light classifier,then use smoothing and highlighting method to process the binarized median saliency map.Finally,we can obtain precise and complete saliency detection result.Different from the salient object detection based on single image,cosalient object detection based on image group requires to detect the same or similar salient objects in two or more images.Detected common objects not only need to be salient,but also to appear in more than half of the images.Considering the characteristics of co-salient object detection,exemplar based co-saliency detection can also be implemented when we take foreground of co-images to generate exemplars and train them with one-class classifier.Extensive experiment shows promising performance of our saliency and co-saliency method.