Research On Key Techniques Of Image Thresholding

With the development of computer science and technology, image processing andcoputer vision techniques have increasingly developted as well, which have been applied tothe actual production and living such as medical image processing, quality inspection inindustrical production, objects locating and tracking in military field, video tracking andmonitoring in intelligent transportation systems, text segmentation and recognition of theprecious historical documents, image and video editing in practice, etc. Image segmentationtechnique, which is a foundation of the image processing techniques and the computer visiontechniques, plays a crucial role in the aforementioned applications. The sequent operation,such as objects locating, objects recognition, objects tracking, image understanding, sceneanalysis and so on, depends intensively on the quality of the segmentation.So far a detailed research for image segmentation techniques has been carried out and agreat of achievements have been made, i.e. many image segmentation algorithms have beenproposed. According to the difference of the image type, the image segmentation algorithmscan be divided into gray image and color image segmentation. According to that whether theuser interaction is included or not in the segmentation procedure, it can be divided intounsupervised (or automatic) and supervised (or interactive) segmentation. In terms of thedifferent image representation it can be divided into pixel based image and superpixel basedimage segmentation, as well as single scale and multi-scale segmentation. In terms of theimage attributes used in the segmentation it can be divided into single attribute andmulti-attribute segmentation. According to the different operation space it can be divided intofeature based and spatially guided segmentation. According to the difference of the driver theimage segmentation can be divided into edge base and region based segmentation. Among theimage segmentation algorithms above mentioned, in this thesis we make a deep research forthe thresholding technique due to characteristics such as real time, simplicity, effectiveness,automation, wide application etc. The detail of the description is as follows.1. A threshold optimization framework of global thresholding algorithms using Gaussianfitting is proposed with the consideration of that those thresholds derived from Otsu method,maximum Entropy method and minimum error method are not the optimal thresholds at all.Firstly, a global thresholding approach is applied to the gray image and dividing it into background and object two parts roughly. And then, two Gaussian populations are fittedaccording to the mean and variance of every part. Since the optimal threshold of two Gaussiandistribution is at the intersection of them, the presented framework optimizes thresholds usingiterative approach until converging to the optimal threshold position. In order to improve theanti-noise performance, a robust threshold optimization framework of global thresholdingalgorithms using Gaussian fitting is introduced combining with Otsu thresholding algorithmbased on rebuilding and dimension reduction of the3-dimensional histogram. Finally,extensive experiments are performed and the results show that those thresholds derived fromOtsu scheme, maximum entropy scheme and minimum error scheme using the proposedthreshold optimization framework can be converge to the optimal threshold position. Plus, thepresented algorithm has robust anti-noise performance and high-efficiency.2. A robust minimum error thresholding method is proposed combiningthree-dimensional (3-D) minimum error thresholding scheme based on2-D method with theprinciple of rebuilding and dimension reduction of the3-D histogram. Taking into account theglobal behavior of this appraoch and the limitation to processing even illumination imagesonly, a Water Flow model is used to estimate background of uneven illumination images forimproving adaptability of the proposed method. And then difference image between originalimage and background can be easily obtained to reduce interference of uneven illuminationduring the binarization process. To improve execution performance of segmentationprocedure, gamma correction is employed to enhance image, accompanying with a globalsegmentation using robust minimum error thresholding algorithm. Subsequently, imagesegmentation tests are carried out with even and uneven illumination, and then comparisonson misclassification error and time expenditure are performed between the proposed methodand other approaches, e.g.1-D/2-D minimum error thresholding, Otsu thresholding algorithmbased on3-D histogram rebuilding and dimensionality reduction, adaptive gray wavetransformation thresholding scheme, as well as bias field based FCM method. The resultsshow that the introduced approach yields better thresholding performance than these methods.3. An adaptive image thresholding algorithm by background estimation in Gaussian scalespace is proposed for thresholding images with uneven illumination. Firstly, a Gaussian scalespace, which is produced from the convolution of a two-dimensional Gaussian function withan input image, is used to estimate the background image. Followed by backgroundsubtraction, the objective image can be easily obtained to eliminate interference of unevenillumination. Secondly, to highlight those darker objects, gamma correction is employed toenhance the objective image. Finally, the thresholding result is extracted easily using the global valley-emphasis Otsu method. To test the effectiveness of the introduced scheme,image segmentation tests are carried out for document and non-document images with unevenillumination, and then comparisons on misclassification error (ME) and time expenditure areperformed among the proposed approach, the biased field-based FCM method, the adaptivegray wave transformation thresholding scheme and the adaptive minimum error thresholdingalgorithm. The experimental results show that the introduced method yields better visualquality and lower ME values than these three approaches.4. A histogram-based color image fuzzy clustering algorithm is proposed for addressingthe problem of low efficiency due to computational complexity and poor clusteringperformance. Firstly, the presented scheme constructs the red, the green and the bluecomponent histograms of a given color image respectively. To keeping smoothness of eachcomponent histogram, pre-processing is employed to each of them. Secondly, the proposedalgorithm multi-thresholds each component histogram using some dominating valleysidentified from a fast peak-valley location scheme in each global histogram. Thirdly,reconstructing a new histogram applying a histogram merging scheme to RGB threecomponent histograms and multi-thresholding this new histogram again using somedominating valleys obtained from the fast peak-valley location scheme, so the proposedapproach can easily identify the initialization condition of cluster centroids and centroidnumber. Finally, we construct a new dataset composed of some pre-segmented small regionsusing the WaterShed algorithm and the FCM algorithm is performed on this dataset instead ofpixels combined with the initial cluster centroids. Experimental results have demonstrated thatthe proposed algorithm is more efficient than the DSRPCL algorithm and the HTFCMalgorithm in the running times and the PRI values.In this thesis through the detail study of the thresholding algorithms, we effectivelyimprove the anti-noise performance and the suppression performance to uneven illuminationfor the thresholding algorithms. And most important the overall segmentation performance ofthe proposed thresholding alrorithms is much better than those original ones. In this work wemainly apply the thresholding technique to the industrial image segmentation for theautomation in the industrial field. In addition we also apply the the thresholding technique tothe color image clustering, which provides research foundation and technical support for thosecomputer vision tasks such as object segmentation, locating and recognition.