Optimum Edge Detection for Object-Background Picture

Optimizing Edge Detectors for Robust Automatic Threshold Selection: Coping with Edge Curvature and Noise

“…If we have a weight function w(x, y) which has all required properties, it automatically follows that w n (x, y) is a suitable weight function too, for all positive n. However, if w(x, y) has a sharp global maximum at the step, w n (x, y) should peak even more sharply around the step, and may therefore have a lower effective radius r eff . This has been pointed out by Kammoun and Astruc [4], who explored the one-dimensional case with asymmetric blurring of the edge. By peaking more strongly around the inflection point of the luminance, T shifted away from B + A/2, and nearer to the grey level of the inflection point.…”

“…Kammoun and Astruc [4] also use higher order filters to reduce the noise bias in the segmentation method. However, higher order filters may also show more sensitivity to random noise, since outliers in the noise distribution are multiplied in a similar way to object related gradients.…”

“…The weight assigned to each pixel is the response of a simple edge detector at that pixel. It has been suggested that many edge detectors other than the original one for which Kittler et al demonstrated the correctness of their approach may be used instead [3,4]. Criteria for optimization of the edge detector are developed from a simple proof of this assumption.…”

“…Bias caused by either uniform Gaussian or Poison noise is discussed. An approach to eliminate such a bias discussed by Kittler et al [2] is extended to the wider class of generalized RATS threshold selectors and compared to the use of higher order filters as suggested by Kammoun and Astruc [4]. Finally, a number of edge detectors are compared in curvature bias, noise performance, and sensitivity, both in theory and on computer-generated images.…”