The ability to automatically find objects of interest in images is useful in the areas of compression, indexing and retrieval, re-targeting, and so on. There are two classes of such algorithms – those that find any object of interest with no prior knowledge, independent of the task, and those that find specific objects of interest known a priori. The former class of algorithms tries to detect objects in images that stand-out, i.e. are salient, by virtue of being different from the rest of the image and consequently capture our attention. The detection is generic in this case as there is no specific object we are trying to locate. The latter class of algorithms detects specific known objects of interest and often requires training using features extracted from known examples. In this thesis we address various aspects of finding objects of interest under the topics of saliency detection and object detection. We present two saliency detection algorithms that rely on the principle of center-surround contrast. These two algorithms are shown to be superior to several state-of-the-art techniques in terms of precision and recall measures with respect to a ground truth. They output full-resolution saliency maps, are simpler to implement, and are computationally more efficient than most existing algorithms. We further establish the relevance of our saliency detection algorithms by using them for the known applications of object segmentation and image re-targeting. We first present three different techniques for salient object segmentation using our saliency maps that are based on clustering, graph-cuts, and geodesic distance based labeling. We then demonstrate the use of our saliency maps for a popular technique of content-aware image resizing and compare the result with that of existing methods. Our saliency maps prove to be a much more effective replacement for conventional gradient maps for providing automatic content-awareness. Just as it is important to find regions of interest in images, it is also important to find interesting images within a large collection of images. We therefore extend the notion of saliency detection in images to image databases. We propose an algorithm for finding salient images in a database. Apart from finding such images we also present two novel techniques for creating visually appealing summaries in the form of collages and mosaics. Finally, we address the problem of finding specific known objects of interest in images. Specifically, we deal with the feature extraction step that is a pre-requisite for any technique in this domain. In this context, we first present a superpixel segmentation algorithm that outperforms previous algorithms in terms quantitative measures of under-segmentation error and boundary recall. Our superpixel segmentation algorithm also offers several other advantages over existing algorithms like compactness, uniform size, control on the number of superpixels, and computational efficiency. We prove the effectiveness of our superpixels by deploying them in existing algorithms, specifically, an object class detection technique and a graph based algorithm, and improving their performance. We also present the result of using our superpixels in a technique for detecting mitochondria in noisy medical images.
EPFL2011
