A Study And Application Of New Methods On Image Filtering And Image Segmentation

Image filtering and image segmentation are key issues in image processing,computer vision, and pattern recognition. With the wide use of image processingtechniques in many fields of production and consumption, great economic and socialbenefits have been brought to us. In recent years, the image processing techniques areincreasingly wide and deep used, and cannot be prevented from expanding. This paperis concerned with the technique of image filtering and image segmentation. In thecontext, we propose some new methods on image filtering and image segmentation,as well as apply all the methods on practical optical interference patterns and medicalimages. It mainly includes the work as follow.A modified spin filter is presented. This method has introduced a newdetermination about the fringe tangent direction and a noise detecting technique at thetangent direction, so it is capable of filtering out the noise with less blurring anddistorting affection on the fringes as well as restraining the isolated noisy points inhigh density station, and it becomes a powerful tool for ESPI images smoothing.An edge detector based on the idea of the modified spin filter is presented. Thenew algorithm uses gray/color intensity change values on the two sides of the currenttangent line to replace the gray/color intensity values of each pixel as to generate anedge map. For the noisy images, the new detector extracts the edges directly withoutany smoothing preprocess.A new neural network model, which is called an enhancing learning on the radialbasis function (ERBF) neural network, is presented to perform color imagesegmentation. ERBF neural network first employs a dynamic algorithm to set its fronttwo layers, the input and the hidden layer. And then it introduces the Hebb rule totrain the hidden layer in order to divide its hidden neurons center vectors into twomeaningful groups: one group members are the object color-clustering centers;theothers are the background centers. Finally, it has the output layer trained by thecompetitive algorithm and puts out different values with different input values, hencedivides the object pixels from the image. The new network model needs not to begiven the nodes number beforehand without troubles in choosing suitable nodesnumbers ahead. And it needs to be trained only twice, but not many times, avoidingsuch a fussy training process as those classical models.