An Efficient Parallel Algorithm for Multi-Scale Analysis of Connected Components in Gigapixel Images

Abstract

Differential Morphological Profiles (DMPs) and their generalized Differential Attribute Profiles (DAPs) are spatial signatures used in the classification of earth observation data. The Characteristic-Salience-Leveling (CSL) is a model allowing the compression and storage of the multi-scale information contained in the DMPs and DAPs into raster data layers, used for further analytic purposes. Computing DMPs or DAPs is often constrained by the size of the input data and scene complexity. Addressing very high resolution remote sensing gigascale images, this paper presents a new concurrent algorithm based on the Max-Tree structure that allows the efficient computation of CSL. The algorithm extends the "one-pass" method for computation of DAPs, and delivers an attribute zone segmentation of the underlying trees. The DAP vector field and the set of multi-scale characteristics are computed separately and in a similar fashion to concurrent attribute filters. Experiments on test images of 3.48 to 3.96 Gpixel showed an average computational speed of 59.85 Mpixel per second, or 3.59 Gpixel per minute on a single 2U rack server with 64 opteron cores. The new algorithms could be extended to morphological keypoint detectors capable of handling gigascale images.

Distributed Connected Component Filtering and Analysis in 2D and 3D Tera-Scale Data Sets

“…One of these tools is the pattern spectrum [15], obtained by applying a set of consecutive filters and extracting the subset of image content that is removed at each step. An other approach is the differential attribute profile (DAP) [16] and its compact representation known as the Characteristic-Salience-Leveling (CSL) model [2]. In [12], we described an efficient implementation of these tools for distributed tree representations.…”

“…Component trees [1] are powerful hierarchical structures representing all connected components at all threshold sets of an image. They provide a suitable approach for region-based analysis, including various filtering strategies and multi-scale tools [2]. They have been applied in astronomy [3], [4], [5], remote-sensing [6], [7], or medical analysis [8], [9].…”

Distributed Component Forests in 2-D: Hierarchical Image Representations Suitable for Tera-Scale Images

“…During the¯nal merge, carried out by the thread with rank zero, the entire component tree and image must be accessible to that thread. Once the¯nal merge has been performed, each thread can proceed to¯lter the image independently, 29 or perform more complex operations like computation of di®erential attribute pro¯les, 31 or pattern spectra. 30…”

“…26 These tree structures form a compact representation of all the connected components of all threshold sets in an image. They¯nd use in various applications, such as attributē ltering, 3,13,20,23 computation of pattern spectra, 24 or morphological pro¯les 1,18,31 and visualization. 27 The computational complexity of algorithms to construct these tree structures is also modest: either OðGNÞ, with G the number of gray levels in the regular 8-16 bit per pixel case or ðN log NÞ in the 32-bit or more integer or°oating point case.…”

“…2,15 Furthermore, shared-memory parallel algorithms for computation of these trees, 14,29 and subsequent postprocessing have been developed. 30,31 On fairly modest compute servers they can handle images up to a few gigapixel at most. This is despite the attractive property of these tree structures, that they form a very memory-e±cient multi-scale representation of an image, which can represent all scales present in the image, unlike classical image pyramids, or wavelets, which discretize scale in some way.…”