Description
An image segmentation process is one of the most important steps in an image recognition or analysis application pipeline. It is a step that splits each image into disjointed regions of interest. It is also a task that is usually performed by biological processes, such as human visual system. Due to the low processing and ease of implementation, one of the most used techniques is the thresholding method, which consists in finding the best cutting thresholds of a probability distribution histogram. However, the higher the number of thresholds, the greater the computational complexity. And there is no consensus on the number of thresholds and the partitioning position as well. This paper presents a study of the number of thresholds for segmenting an image into their regions of interest. For this purpose, the proposed method uses a bio-inspired algorithm based on meta-heuristics, called firefly with a non-extensive Tsallis statistics kernel. Also, the images are pre-filtered with a low-pass filter based on a q-gaussian function. Using a manually segmented database, the results show that there is an inverse correlation between the Fourier spectrum of an image and the number of thresholds which most approximates the image from the used ground truth. This suggests an automatic method for calculating the required number of thresholds.
