Sergey Zavalishin

Fast Algorithm for Visibility Enhancement of the Images with Low Local Contrast

In the paper, we propose new fast and effective approach for automatic visibility enhancement of images with poor global and local contrast. Initially, we developed the technique for scanned images with dark and light background regions and low visibility of foreground objects in both types of regions. Newly proposed algorithm carries out locally adaptive tone mapping by means of variable S-shaped curve. We use cubic Hermit spline. Starting and ending points of the spline depend on global brightness contrast, whereas tangents depend on local distribution of background and foreground pixels. Alteration of the tangents for adjacent areas is smoothed in order to avoid forming of visible artifacts. The description of several optimization tricks, which allow realization of high-speed algorithm, is given. We compare proposed method with several well-known image enhancement techniques by means of estimation of Michelson contrast (also known as visibility metric) for a number of test patterns. Disclosed algorithm outperforms tested alternatives. Finally, we extend application of proposed method for photo enhancement and correction of images with haze.

Parameters adaptation framework for local contrast enhancement

In this paper, we propose a novel local contrast enhancement algorithm based on preliminary image segmentation. Our algorithm adjusts contrast to individual image regions and applies further weighed least squares smoothing in order to reproduce brightness transitions similar to original image. We also measure region similarity using color and size features and apply the same contrasting parameters to similar regions. Proposed approach performs dramatic contrast improvement along with both dark and light details preservation. Comparison with famous algorithms demonstrates that our algorithm outperforms them according to several contrast metrics.

Document Image Classification on the Basis of Layout Information

In this paper, we propose a document image classification framework based on layout information. Our method does not use OCR; hence, it is completely language independent. Still we are able to exploit text data by extracting text regions with a novel MSER-based approach. Our MSER formulation provides great robustness against text distortions in comparison to the existing one. We introduce two types of novel image descriptors supplemented with Fisher vectors, based on Bernoulli mixture model. Classifiers, based on aforementioned descriptors, are assembled into meta-classification system that is able to classify document in complex cases when individual classifier accuracy is poor. Our meta-classification system demonstrates low processing time comparable to a single classifier. We show that our method outperforms the existing ones by the means of classification accuracy for a wide range of documents of both well-known and machine-generated document datasets.

Inverse halftonning using sparse coding methods

We propose a new inverse halftonning algorithm based on sparse coding approach, which is able to simultaneously reconstruct the continuous-tone image from the halftonned one and increase spatial image resolution. To achieve this goal, we introduce the halftonning operator that maps the continuous-tone signal into the same representation as the halftonned one. Resolution enhancement is obtained with a pair of the linked dictionaries which consist of low resolution and high resolution patches. Comparison with some famous algorithms has demonstrated that our algorithm outperformed them by both image quality and applicability to different halftone types.

A run equivalence algorithm for parallel connected component labeling on CPU

It is proposed a new algorithm for parallel connected component labeling which applies labels to image runs. The algorithm is designed to efficiently label a text document with a large number of characters inside. In contrast to the existing parallel labeling algorithms, our method benefits from modern CPU architectures; it is designed to operate on image runs, not pixels, which make it possible to minimize a number of memory read-write operations. Each CPU core processes a bunch of runs, aligned by rows, that has a positive impact to CPU and memory cache utilization. The results of modeling have shown that our algorithm demonstrates better performance than existing CPU-based connected component labeling algorithms. Moreover, the developed algorithm also demonstrates a good scalability across different numbers of CPU cores.