Research On Wavelet-based Edge Detection And Edge Line Features Description Algorithms

Edge is one of the important features which carrying most information of theimage. Edge detection and edge line characterization representation are the focus ofcomputer vision research and the foundations for higher level of image processing.Good edge detector should be high accuracy and robust to noise. However, accuracyand noise immunity are incompatible, all the algorithms can not meet high locationaccuracy and strong resistance to noise at the same time up to now. For the influenceof complex image environment and the effect of edge detection, most of image edgesshow irregular geometry. The edge line features is difficult to be described byconventional geometry method. The edge line feature descriptors not only should havegood translation, rotation and scale invariance, but also should have able to inhibit theinterference of noise with good stability. Most of the existing methods can not meetthese requirements.Wavelet transform has good performance at multi-resolution analysis in timedomain and frequency domain. It is very suitable for image edge detection. Thedirection information of the pixels in the image has not been fully utilized fortraditional edge detection methods based on wavelet. Detection performance isdissatisfactory when images are affected by noise seriously or the contrast is low. Inthis paper, we proposed a3thresholds edge detection algorithm based on multi-scalefusion in wavelet domain, the advantage of this novel method is the excellentanti-noise performance. We proposed the concept of gradient direction characteristicsutilizing the differences between the edge point and noise point along the gradientdirection. First of all, image is transformed into wavelet domain on different scale toget gradient information for each pixel. Secondly, edge point is detected based on thenon-maximum inhibition method of dual-threshold and the gradient directioncharacteristics. Finally, the detection results in each scale are fused and the edgeextracting results is obtained by the third threshold. Experiment results denominate ouralgorithm not only has good capacity of edge location, but also can extract satisfactoryresults for the image involved with strong noise.Fourier curve description algorithm based on mature Fourier analysis theory is theoptimum algorithm for curve description, its concept is simple and can be understoodeasily, the global and local information of the curve can be reflected at the same time.The existing Fourier description algorithms mostly depend on the sample of curve. The edge description effect is affected seriously by edge detection results and the smallperturbation of curve. To improve the performance of the curve descriptor, animproved algorithm of Fourier contour curve description is proposed in this paper.Firstly, the cycle shift sampling algorithm is used to obtain multiple sets of curvesample points. Secondly, Fourier transform is performed for each point sets. Finally,the minimum set of standardized Fourier coefficients is chosen as the Fourierdescriptor of the curve. Our algorithm is more concise and stable than traditional onefor it avoids the complex sampling process. Experimental results show that ouralgorithm can describe the curve more accurately and can reflect the differencebetween the different curves more effectively compared with the traditional Fouriercurve description algorithm.