Background: In this research, a novel algorithm is formulated through the combination of gradient and adaptive thresholding. A set of 5 X 5 convolution kernels were generated to determine the gradients in the four main directions of the image.
Objectives: The researcher converted the gaussian equation into a normalized kernel, which was convolved with the gradients to suppress the impact of noise.
Methods: The edges derived were partitioned into a set of 5 x 5 matrices. A weighted variance was calculated for each local window in the image. The pixel that generated the minimum variance was used for the segmentation process in each local window. The researcher then trimmed multiple pixel width edges into singles by developing a set of 5 X 5 Structuring Elements (SE). These elements were placed over the image to remove boundary pixels. In order to produce colored edges, the algorithm was executed over all the channels and the results were concatenated to produce the skeletal colored edges.
Results: From the evaluations conducted, the proposed algorithm exhibited better performance than most of the recent algorithms with respect to Human Perception Clarity and time complexity in both noisy and nonuniform illuminated images.
Conclusion: The reason for this performance is that it is able to extract edges moving in the various directions of images. It also ensures that identified edges are single pixel width instead of multiple.
Keywords: Image segmentation, Human Perception Clarity, Weighted variance, Gradient, Skeletonization, Normalization, Edge detection