Noise Processing And Sub-pixel Feature Extraction For Microscopic Images

Micro stereovision system based on stereo light microscope (SLM) has been applied in some micro-domains widely. Highly accurate measurement and location are implemented by the vision feedback. As far as a vision system is concerned, its accuracy is affected mainly by noise, system resolution and geometrical distortion. For these factors, noise processing and sub-pixel feature extraction for SLM microscopic images are studied in this papa.Images are polluted by noise in the processes of imaging and transmission, and result in radiometric distortion. According to the optic or electronic characteristics of each part of SLM micro stereovision system, the noise resources are analysed for SLM microscopic images. Moreover, an image model is built to describe the relationship between image signal and noise. Noise is sorted into random noise and non-uniformity noise according to its characteristics. For random noise, a vector median filter for color images is developed based on the rank of the sum of 1-norm distances and 2-norm distances from one pixel to the other pixels in the filtering window and combining with an average filter. In order to avoid the computational complexity of the vector median filter, a fast algorithm for speeding up vector median filtering is presented. The algorithm includes various effective fast algorithms and can speed up 76.3% ~ 93.4%. The experiment results demonstrate that the fast vector median filter is excellent in noise removal and highly efficient. For non-uniformity noise, two correction methods based on non-uniformity calibration are developed. A vector median filter is adopted to improve the calibration accuracy of the average calibration method based on n flat-field images, and an iterative method is used to mend precision of the fitting calibration method based on one flat-field image. Consequently, two non-uniformity calibration methods for non-uniformity correction are developed. The experiment results show that non-uniformity correction is very necessary for SLM microscopic images, and also demonstrate that the calibration results of the two methods are very similar and both are satisfied. The image quality is improved greatly after non-uniformity correction and filtering.The sub-pixel location technology is a method to improve the resolution of a vision system through software. Considering the characteristics of SLM microscopic images and color information of them, two wavelet transformation based methods in HSI color space and vector space for edge detection are proposed. Then, the interested area of the image is interpolatedusing cubic spline and accurate sub-pixel edges are obtained based on the zero point of first-order derivatives. The experiment results show that the methods are effective for SLM microscopic images with various magnifications. And it is manifested that the experiment results of the two methods are very similar. For images with impulse noise whose pepper and salt proportions are 0 ~ 0.15, the position error is 0 ~ 0.23 pixels. For images with Gaussian noise whose standard deviation is 0 - 60, the position error is 0 ~ 0.16 pixels. As an example of sub-pixel edge detection, a method for extracting highly accurate image features is presented for plane calibration templet with grids. By using sub-pixel edge detection and edge tracking etc., accurate image features are extracted. For images with impulse noise whose pepper and salt proportions are 0 ~ 0.15, the position error is 0.052 ~ 0.11 pixels. For images with Gaussian noise whose standard deviation is 0 ~ 60, the position error is 0.052 ~ 0.086 pixels.