Compactness in superpixel segmentation of digital images using perceptual colour difference measure

Abstract

Digital image segmentation is a thrilling but challenging open problem that has been well researched in the fields of computer vision, and image processing. It has many practical applications like biometric identification, ship detection, building extraction, road marking recognition, deoxyribonucleic acid matching, welding inspection, pedestrian re-identification, object tracking, image editing, pest monitoring, and shopping items recommendation. In recent years, image segmentation has come to rely heavily on superpixel methods to circumvent the computational complexity inherent in pixel processing. The superpixel approach is generally used to group similar pixels into a semantic cluster of fewer pixels to increase the processing speed and simplify computational intricacy. However, the reliance on the existing superpixel based segmentation methods on the Euclidean distance metric as a measure of similarity between two pixels in an image presents an inherent challenge. The Euclidean distance has a real-world advantage because of its assumption of non-uniformity that most image colour distributions generally follow. This assumption states that real data will occupy a small clustered subset of the entire space, but not necessarily distributed evenly in a higherdimensional space. However, since it cannot deal with illumination change in images, it is limited in compactly measuring similarity in the context of an application that complies with the human perception of similarity. The human eyes can recognise similar or irrelevant image colours under the illumination change for which the Euclidean distance does not perform well. This study aimed to investigate the performance of an attribute concurrence influence distance metric on image compactness in a superpixel segmentation algorithm. It is hypothesized that superpixel segmentation based on attribute cooccurrence similarity measure is likely to achieve better results than Euclidean distance in terms of the performance metrics of under segmentation error, achievable segmentation accuracy, compactness, boundary recall, and contour density. Superpixel segmentation experiments were performed using two widely used colour models which are hue, saturation, value (HSV), and lightness, redness, yellowness (LAB) with the strong attribute concurrence influence distance (SAID) and Euclidean distance in a superpixel segmentation algorithm. The results presented for the LAB colour model showed that SAID outperformed the Euclidean distance for images reflecting overlapping and complex objects with regular compactness. However, the Euclidean distance performed better than the SAID for images with multiple, centre, and low contrast objects with regular compactness across the under segmentation error, achievable segmentation accuracy, boundary recall and contour density performance evaluation metrics. Consequently, for irregular compactness, SAID further outperformed the Euclidean distance for images with overlapping, complex, multiple, Centred and low contrast objects for boundary recall. However, the Euclidean distance performed better than SAID for under segmentation error, achievable segmentation accuracy, and contour density. Furthermore, the compactness performance for SAID and Euclidean distance gave the same compactness value for both regular and irregular compactness. Consequently, based on the analysis of the results for the HSV colour model, it was observed that performances of SAID and Euclidean with regular compactness were at par across all the performance metrics used for images with overlapping, complex, multiple, centre, and low contrast objects. However, the Euclidean distance outperformed SAID with irregular compactness for images with overlapping, complex, multiple, centre, and low contrast objects.