Research On A Multiscale Edge Detection Algorithm Based On Bi-Directional Tracing

Edges represent one kind of low-level image features,and have been widely used in various image analysis and computer vision tasks.A key problem of edge detection is the choose of scale parameters.Under low scales,edges have accurate locations but contain unnecessary details,resulting in a low precision of edge detection.Under high scales,re-dundant information is suppressed but edges could have deformed shapes,leading to a low recall of edge detection.To get better performance of edge detection,image cues can be in-corporated with each other across multiple scales.An investigation of the above-mentioned topics is conducted in this paper.The main contributions and innovations of our work con-tain the following three parts.The scale-space behaviors of edges are studied and a systematic theory is establish.As the scale value varies continuously,an edge may dislocate,disappear,or merge with another edge.Based on a set of typical edge models,a quantitative derivation of the scale-space trajectory of edges is presented,yielding a series of scale-space properties of image edges.Simulation experiments show that there is a consistence between theoretical edge models and numerical edge models.The scale-space behaviors of complex edge models are also summarized.According to the scale-space theory of edges established above,a multi-scale edge detection algorithm based on a bi-directional tracing strategy is proposed.First,edges are identified at a high scale and then they are traced to the lowest scale for obtaining their accurate locations,so that the edge accuracy is improved.Then,edges are confirmed by tracing them along the low-to-high direction for suppressing those spurious edges.Finally,our multi-scale edge detection algorithm is designed and coded.Experimental results show that the proposed algorithm has better performance subjectively and objectively than the existing edge detection algorithms.A multi-scale edge detection algorithm for color images based on bi-directional tracing is also proposed.The colors of an image contain rich scene information.However,most current edge detection algorithms are designed for gray images.As a result,significant edges could be frequently miss-detected.To overcome this difficulty,the above-established multi-scale edge detection algorithm is extended to process color images.First,a preprocessing step is implemented to convert the gradient vectors of the input image from different color channels into scalars to obtain a gradient map.Then,edges in the gradient map are verified by using a non-maximum suppression.Finally,a bi-directional tracing of the verified edges is conducted in scale space to improve the detection performance.Experimental results show that the new algorithm can solve the problems of the gray edge detection,and it delivers superior performance over the state-of-the-art color edge detection algorithms subjectively and objectively.