Content
- Introduction
- 1. Overview of methods for recognizing objects in images
- 2. Description of bilateral filter
- 3. Directions, goals and objectives of further research
- Findings
- List of sources
Introduction
Computer processing and recognition of objects in images is one of the rapidly developing branches of modern computerization. Recognition of 3D objects from 2D images is one of the most important tasks of scene analysis and computer vision. The basic information for recognition contains images in different parts of the spectrum, various emissions, for example, such as optical, infrared, ultrasound, etc., which were obtained in various ways, such as, for example, television, photographic, laser, radar, radiation, etc., converted to numbers and presented as matrices of numbers.
Studies on the recognition of images of objects are very diverse in the formulation of tasks and the choice of means for their implementation, which is one of the consequences of the diversity of areas of practical application. The basic tasks that were solved even in the first development of computer vision systems are the tasks of detecting and further recognizing objects that have a certain shape based on noisy and uneven images. The pioneer in practical tasks that pushed the formation and development of the industry of recognition of objects in images was the task of identification and recognition of human faces. One of the methods widely used in solving problems of object recognition is the use of a bilateral filter [1,2], but its main significant drawback is the speed of operation. In this regard, it is relevant to consider and study the possibility of parallelizing the filter to increase its quality and performance.
1. Overview of methods for recognizing objects in images
The task of recognizing objects in images is modern and relevant. The need to solve such problems arises in a variety of areas from military affairs and security systems, to the digitization of analog signals. A single recognition algorithm does not exist, therefore, they use certain methods for different subject areas. For example, for optical pattern recognition, you can apply the method of enumerating the appearance of an object at various angles, scales, displacements, etc. For characters, you need to iterate over the font, font properties, etc. Another approach is to find the contour of the object and study its properties (connectivity, the presence of angles). In some areas, neural networks are used to effectively implement pattern recognition, but this method requires either a large number of examples of recognition problems (with correct answers) or a special neural network structure that takes into account the specifics of this task.
2. Description of bilateral filter
The bilateral filter is a nonlinear image smoothing filter with preservation of its boundaries. This filter is always used in image processing, in computer graphics and for solving other problems. The filter has received such a wide range of applications because of its features. Firstly, it is simply formalizable and easy to do. In addition, it is not based on iterations, i.e. a moderate result is achieved after just one study, and secondly, the filter is not linear and is not used without any kind of improvement, because It requires quite a lot of computing power, which not every computer can afford. Thus, processing an image of 1024x1024 pixels on a standard desktop computer will be about an hour.
It is worth noting that the algorithms that accelerate the filter were developed earlier, but the vast majority of them result in a loss of quality. Such algorithms describe a set of methods that speed up the filter. At the moment there are two principles of filter acceleration, the first are specialized filters that perform accurate calculations, but are limited in certain tasks. The second principle is to use rounding filters, they are general in nature, but do not allow to achieve over-accurate results. Mathematicians Durand and Dorsey converted this filter to a linear form, which allows the use of fast Fourier transforms. To increase the speed of calculations to one second or less, they use data sampling, which is reduced, for small images. In their work, for a bilateral filter, space is allocated for calculating the fast Fourier transform, and direct convolution becomes more efficient as soon as the data is converted, since the kernel is quite small. Although their study does not highlight this, the results of their work are achieved with direct convolution, without using the Fourier transform. The technique they developed is directly related to their work in that a linear filter is expressed through linear operations and the largest part of the acceleration is taken from downsampling. This allows you to get improved image processing accuracy when using this filtering. A bilateral filter is also used to filter the depth map. The filtered depth map has a satisfactory visual quality because the fast bilateral filter generates a smooth depth map inside a smooth region with similar pixel values and maintains a sharp gap in depth at the boundary of the object. After filtering with a quick filter, the depth map is then used to generate left, right, or multi-image viewing using image-based depth rendering (DIBR) for 3D rendering.
3. Directions, goals and objectives of further research
The analysis showed that as the preferred method for solving the problem of accelerating bilateral filtering, you can choose the method of decomposition of the bilateral filter into spatial filters, because it can process images with arbitrary spatial and range cores, which increases the ability to process images of different sizes and with different contents.
This method can also be fairly easily implemented in parallel. For example, on an NVIDIA Geforce 8800 GPU with the same quality As a result, the method is on average 10 times faster than other methods [8]. Since the method uses independent processing of each pixel, the method can be parallelized, which will speed up its work, and it is possible to carry out image processing in real time.
The purpose and objectives of further research is to analyze the functioning of the method and study the characteristics of its work, with further optimization of the method in terms of quality and speed of work, as well as mapping the obtained modified method to parallel structures of computing systems, obtaining the characteristics of the implementations and analyzing the effectiveness of the completed developments. It is planned to use a parallel multiprocessor of a PC video card as a parallel computing system, organize work with it using GPGPU and OpenACC [10] technologies.
Findings
In the course of the work, an analysis was made of existing means of recognizing objects in images, methods and algorithms for bilateral filtering for solving problems of recognizing objects in images were considered. A description of the bilateral filter and its mathematical representation is given, the problems of the filter and methods for solving them are analyzed. The possibilities of accelerating and improving the quality of bilateral filtering during image processing due to parallelization of computational processes were considered, a preliminary analysis of the possibility of parallelizing the filter was performed. The method of decomposition of the bilateral filter into spatial filters and its further parallel implementation was chosen as the object of further research. It is planned to use a parallel multiprocessor of a PC video card as a parallel computing system, organize work with it using GPGPU and OpenACC technologies.
List of sources
- Билатеральные фильтры кратко. [Электронный ресурс] – Режим доступа: http://old.unick-soft.ru – Загл. с экрана.
- Bilateral filter [Электронный ресурс] – Режим доступа: https://en.wikipedia.org – Загл. с экрана.
- Фурман Я.А Введение в контурный анализ. Москва, ФИЗМАТЛИТ, 2003, 598 с.
- Обзор основных методов контурного анализа для выделения контуров движущихся объектов [Электронный ресурс] – Режим доступа: http://www.engjournal.ru – Загл. с экрана.
- Weickert J., Coherence-Encahncing Shock Filters. Lecture Notes in Computer Science. Springer, 2003, vol. 2781, pp. 1–8.
- Метод предварительной фильтрации изображений для повышения точности распознавания образов. [Электронный ресурс] – Режим доступа: http://www.engjournal.ru – Загл. с экрана.
- W.-Y. Chen and Y.-L. Chang and S.-F. Lin and L.-F. Ding and L.-G. Chen.”Efficient depth image based rendering with edge dependent depth filter and interpolation,” in Proc. ICME, pp. 1314-1317, 2005.
- Анализ эффективности алгоритмов билатеральной фильтрации. [Электронный ресурс] – Режим доступа: http://engineering-science.ru – Загл. с экрана.
- Real-Time Bilateral Filtering. [Электронный ресурс] – Режим доступа: http://vision.ai.illinois.edu – Загл. с экрана.
- OpenACC. [Электронный ресурс] – Режим доступа: https://www.openacc.org – Загл. с экрана.