Optimized Connected Components Labeling with Pixel Prediction

Abstract

In this paper we propose a new paradigm for connected components labeling, which employs a general approach to minimize the number of memory accesses, by exploiting the information provided by already seen pixels, removing the need to check them again. The scan phase of our proposed algorithm is ruled by a forest of decision trees connected into a single graph. Every tree derives from a reduction of the complete optimal decision tree. Experimental results demonstrated that on low density images our method is slightly faster than the fastest conventional labeling algorithms

Connected Components Labeling on DRAGs

“…In the following, we will use acronyms to refer to the compared algorithms: CT is the Contour Tracing approach by Fu Chang et al [6], SAUF is the Scan Array Union Find algorithm by Wu et al [15], BBDT is the Block Based with Decision Trees algorithm by Grana et al [2], CTB is the Configuration-Transition-Based algorithm by He et al [17], and PRED is the Optimized Pixel Prediction by Grana et al [18]. Moreover, labeling NULL is a lower bound limit for all CCL algorithms, obtained by reading once the input image and writing it on the output again [22].…”

“…In 2014, He et al [17] introduced the Configuration Transition Based (CTB) algorithm in which they observed that during the first scan some checks can be avoided if they have been already performed in the previous step. Following this approach, Grana et al in [18] proved a general paradigm to exploit already seen pixels during the scan phase, in order to minimize the number of times a pixel is accessed. In fact, the same decision tree is usually traversed for each pixel of the input image, without exploiting values seen in the previous iteration.…”

Spaghetti Labeling: Directed Acyclic Graphs for Block-Based Connected Components Labeling

“…Grana et al [25], instead, proposed a general paradigm to leverage already seen pixels, which combines configuration transitions with the decision trees. Algorithms that make use of a DTree usually traverse the same tree for each pixel of the input image.…”

“…-a proved algorithm to produce optimal decision trees [24]; -realizing that it is possible to use a finite state machine to summarize the value of pixels already inspected by the horizontally moving scan mask [16]; -combining decision trees and configuration transitions in a decision forest, in which each previous pattern allows to "predict" some of the current configuration pixels values, thus allowing automatic code generation [25]; -switching from decision trees to Directed Rooted Acyclic Graphs (DRAGs), to reduce the machine code footprint and lessen its impact on the instruction cache [26].…”

“…On the other hand, the procedure proposed in [25] is suitable to be automatized, but still a small mask was employed. The reason, in this case, is that the larger the mask is, the more decision trees will populate the resulting forest, and the higher every tree will be.…”