Learning Visual Dictionaries from Class-Specific Superpixel Segmentation

Superpixel Segmentation using Dynamic and Iterative Spanning Forest

“…It is worth noting that DISF cannot provide a hierarchical segmentation, but its strategy to select seeds based on relevance and its path-cost function based on dynamic arc-weight estimation can be explored in superpixel graphs for hierarchical segmentation [9]. Similarly, its higher effectiveness in delineation might better define symmetrical supervoxels for brain asymmetry analysis [15] and class-specific superpixels for image description [6]. We also intend to investigate extensions of DISF that incorporate prior object information in the path-cost function [3].…”

“…In summary, our contributions are: (1) a new three-step pipeline for seed-based superpixel segmentation, which aims to include relevant seeds in the initial seed set and retain them during the process; (2) a new ISF-based method, named DISF, which relies on the new pipeline; (3) rules to estimate seed relevance and number of irrelevant seeds for removal at each iteration, such that the desired number of superpixels is always achieved at the end of the process; and (4) the incorporation of dynamic arc-weight estimation in IFT-based superpixel delineation for more effective segmentation. Some of these contributions can also benefit ISF-based methods [3], [9], [15], [6] recently developed for distinct applications.…”

Intestinal Parasites Classification Using Deep Belief Networks

“…Three different models were considered in this paper: one RBM with 500 hidden neurons and two DBNs, i.e., the first with two hidden layers (DBN-2) containing 500 neurons each, and the other comprising three hidden layers (DBN-3) with 2, 000 neurons in the first two levels and 500 neurons in the uppermost layer 3 . All models were trained for 100 epochs considering each RBM stack with a learning rate η = 10 −5 and mini-batches of 64 samples.…”

“…Furthermore, Benato et al [1] investigated an approach to cope with the lack of supervised data by interactively propagating labels to reduce the user effort in data annotation. Finally, Castelo et al [3] used bag of visual words to extract key points from superpixel-segmented images and further build a visual dictionary to automatic classify intestinal parasites.…”

Efficient Multiscale Object-based Superpixel Framework

“…The Image Foresting Transform (IFT) [14] is an elegant framework for developing connectivity-based image operators, and can be computed using a generalization of the Dijkstra's algorithm. Due to its object delineation performance, it has been used in several applications [23,29,11], especially for superpixel segmentation [33,8,15]. In this work, we consider its seed-restricted variant.…”