Research On Edge Detection And Its Evaluation

Edge detections are the main research hotspots and play an important role in image processing, computer vision, and pattern recognition. Various edge detection and evaluation algorithms have been proposed in the past decades, but so far there is neither a unified edge detection algorithm nor a widely accepted evaluation technique. There is necessity to investigate more efficient method in this area. This dissertation researches on edge detection and detection results evaluation, and includes the following major contributions:1. A quantitative evaluation index for connectivity of edge is proposed. Firstly, the contributions of edge pixels to the connectivity of edge segments are measured by distances and the edge scale. Then, a monotonous increasing function is constructed to calculate the connectivity indexes of edge segments. The weighting mean value of these connectivity indexes is taken as the connectivity evaluation index of an edge image. The results of horizontal and longitudinal comparative experiments show that this index can evaluate objectively the connectivity of an edge image, and lays the foundation for effectively evaluating edge images.2. An edge image reconstruction method is proposed. This method searches the original pixels for reconstruction from different angles, and then reconstructs the new pixel by linear interpolation which is weighted by the psychological distances between the new and the original pixels. Compared with the existing edge-based image reconstruction methods, this proposed method can overcome the problems of region distortion and the occurrence of false edges, and have higher similarity between the reconstructed and the original images.3. An unsupervised method for edge detection evaluation is presented based on the reconstructed similarity,the confidence level, and the connectivity evaluation index. The reconstructed similarity measures the integrity and the localization veracity of edges; the confidence level reflects the proportion of real and false edge pixels; and the connectivity evaluation index evaluates the connection of edges. Experimental results indicate that this method can evaluate the edge quality effectively and corresponds with human intuitionistic evaluation closely.4. An unsupervised threshold selection method for edge detection is proposed to use of edge pixels spatial relationships and reduce false edges. A series of edge images from a set of descending-order threshold candidates are generated at first, and then their connectivity evaluation indexes and the number of continuous edge segmentions are computed. Finally, the selected threshold is determined by the maximum that the corresponding connectivity evaluation index reaches and the continuous edge segmention on the right of the images increases greatly. Compared with the existing Otsu, Rosin, and Medina methods, the proposed method has higher stability and reliability.5. An edge detection method is proposed by combining hysteresis linking and forecasting-based searching. The continuous edges are connected according to the change of gradients. In order to detect the broken edges with low gradients, this method uses the local information of the edge pixels to forecast and then searches the potential weak edges. Compared with the traditional hysteresis linking methods, this proposed method can not only reduce the threshold sensitiveness, and can but also detect the meaningful weak edges and restraint the meaningless edges with the same gradient values.