Research On Thresholding Technologies For Noisy And Uneven Illumination Images

Image segmentation is an essential part of computer vision, and plays an important role in the overall performance of computer vision system. The purpose of image segmentation is to divide the image into a set of areas that are not overlapped. These areas are either meaningful for the current task, or help to explain the correspondence between them and the actual objects or some parts of the objects. Image segmentation can simplify or change the representation of images, making images easier to be understood and analyzed. It is usually used to locate the objects or edges in the image. More specifically, image segmentation is a process of tagging each pixel, and this process makes the pixels with the same label have a certain common visual feature.Image segmentation has experienced several decades of development so far, and there have been a lot of different types of algorithms. The current image segmentation methods can be generally categorized into: thresholding methods, region-based methods, edge-based methods, the methods combining edge and region, and the multi-scale methods. In addition, neural networks, digital morphology, fuzzy theories are widely applied in image segmentation in recent years, giving birth to many new algorithms.The binarization is a special type of image segmentation, whose purpose is to extract the target (foreground) from the background, and as the result, the target and the background will be labeled with different colors, which are usually black and white. The binarization will provide a basis for post-process steps of feature extraction and scene analysis. It is usually applied in industrial non-destructive detection and character recognition. The common characteristics of the images in these fields are the relatively simple composition of the foreground and background, that rarely containing complex texture. Therefore, the pertinence of the binarization for this type of images is relatively strong.In a variety of binarization algorithms, as a type of conventional algorithms, the thresholding segmentation algorithms have always been widely welcomed for their simplicity and efficiency. They have been researched and developed for decades by many domestic and international researchers, giving birth to a number of outstanding algorithms, which can be categorized into: histogram shape-based methods, clustering-based methods, entropy-based methods, object attribute-based methods, spatial methods, local neighborhood-based methods, and block-based method.However, the thresholding segmentation is still facing many difficult problems to solve currently, such as the adaptability problem, the universality problem, the noise immunity problem, the degradation problem of document image, the non-uniform illumination interference problem, and et.al. In some harsh industrial environments, the images captured for detection are usually subject to uneven illumination or noise and other factors, which seriously affect the accuracy of subsequent recognition stage. Currently, the general approaches for the binarization of industrial inspection image are mainly the global adaptive algorithms and local block-based algorithms. However, the global algorithms is unable to eliminate the uneven illumination. And the lack of instructions in the block division in the local block-based algorithms seriously affects the segmentation quality. Furthermore, the"ghost area"problem is also a difficulty for local block-based algorithms. Therefore, a robust algorithm that is able to effectively resist the interference from uneven illumination and noises is desired for the binarization of industrial inspection images. Meanwhile, sometimes the real-time requirements are high in some industrial inspections, making the algorithm complexity become an important factor for the overall system performance.This paper is focusing on the subject,"The Research on Thresholding Technologies for Noisy and uneven illumination Images", in order to improve the segmentation performance for noisy and uneven illumination images. To solve several current problems in the field of thresholding segmentation, this paper has proposed two corresponding improved algorithms of two existing global adaptive methods, and has introduced a novel concept, namely Gray-intensity Wave Transformation, to weaken the interference from uneven illumination on the industrial inspection images. The main contents of this paper can be listed as follows:1. This paper has presented a three-dimensional Renyi entropy thresholding algorithm based on weighted middle value. This algorithm adds the weighted middle value into the intensity-mean pair in prevent two-dimensional histogram as the third spatial feature. The new three-dimensional histogram could be constructed based on this new triple vector. The thresholding algorithm focus on searching for the best threshold in the three-dimensional histogram that makes the sum of the entropy in background and foreground maximum, in order to threshold the image with a higher noise-immunity than two-dimensional Renyi entropy algorithm. Although the computational complexity of our algorithm is higher than the two-dimensional algorithm, and is even slight higher than some other three-dimensional algorithm, in many actual environments that do not require a high speed, the time-consuming of our algorithm is still tolerable. The performance of the experiments on a number of images had shown the feasibility of our algorithm.2. This paper has presented an Otsu thresholding algorithm based on the reconstruction and dimensionality reduction of three-dimensional algorithm. In this algorithm, it first analyzes the distribution of noise points in the three-dimensional histogram, and then it adjusts the location of each noise point according to the characteristic of the specific type of noise points. Based on these adjustments, a new three-dimensional histogram is reconstructed, which could effectively reduce the noise interference. Moreover, the high dimensionality in the searching process in original three-dimensional Otsu algorithm makes the time complexity of original algorithm too large, which greatly reduces the usefulness of the algorithm. Focusing on this shortage, our algorithm improved the region division in the three-dimensional histogram from eight-partitions to two-partitions, and substituted the triad in the three-dimensional histogram with the triad's vertical distance between the origin and the plane that the triad belonging to. This approach reduces the spatial dimensions of searching threshold from three-dimension to one-dimension, which significantly lowers down the processing time and storage memory. Finally, the visual qualities of several widely used algorithms in the similar type are compared with our algorithm, including 2D Minimum Error method, 2D Renyi Entropy method, 3D Otsu method, etc. And the comparison on time-consuming is also conducted among these algorithms. The experiments had shown that, compared with 3D Otsu algorithm, our algorithm has a smaller time complexity, stronger noise immunity and better visual quality.3. This paper has proposed a novel concept, namely gray-intensity wave transformation, in order to reduce the uneven illumination interference in the segmentation of industrial inspection images. The algorithm views the raw image as a three-dimensional terrain of the intensities, and the uneven illumination has changed the topography of the terrain, i.e. pulling up or down the original waves in the local region. The purpose of gray-intensity wave transformation is to normalize the local terrain affected by uneven illumination into a uniform plane. The wave position of a pixel in 2D intensity wave can be approximated by the positions in the several 1D intensity waves in multiple directions. Therefore, the intensity wave transformation firstly performs the one-dimensional transformation in the specific direction, and then combines the transformed value of the pixel in each direction into a vector, which can approximately represent the 2D wave position of the pixel. These vectors can form a gray-intensity wave transformed matrix. In order to reduce the dimensions of the transformed vector, the Principle Component Analysis method is employed to reduce the vector to one-dimension. This procedure can preserve most of the information in the matrix, and as the result, the gray-intensity wave transformed image is generated, in which the uneven illumination will be weakened. In the post-process step, the classic Otsu algorithm is chosen from a number of global adaptive thresholding algorithms to binarize the transformed image. Only two parameters are needed in our algorithm, namely wave height threshold and Boolean background color. The experimental results have shown that, for the binarization of industrial inspection image under uneven illumination, our algorithm yields better visual quality than several known local adaptive methods. Furthermore, the sensitivity of our algorithm on the key parameter, namely wave height threshold, has also been tested through the experiments. The experimental results have proved that the sensitivity of our algorithm on this parameter is not high.In summary, the paper has studied the thresholding algorithms for noisy and uneven illumination images. Two classic global adaptive thresholding methods have been improved to enhance the visual quality and robustness of the binarization. Furthermore, for the industrial images under the interference of non-uniform illumination, a novel concept of gray-intensity wave transformation has introduced in this paper. The algorithm based on this concept has been proved to be able to effectively reduce the interference of the uneven illumination in the binarization. In a word, this paper has achieved some significant results in the aspects of theoretical exploration and algorithmic application, which would have a positive effect in promoting the application of the image binarization technologies in the future.