IMPLEMENTATION OF THE WATERSHED METHOD FOR SEGMENTING PADANG TRADITIONAL CUISINE IMAGES TO IMPROVE CULINARY OBJECT RECOGNITION ACCURACY

Abstract

This study examines the implementation of the Watershed method for segmenting images of traditional Padang cuisine with the aim of improving culinary object recognition accuracy. Padang dishes possess complex visual characteristics, where multiple food components such as rice, side dishes, chili sauce, and vegetables are presented on a single plate. These elements often overlap and exhibit similar colors and textures, making image segmentation challenging when using conventional methods. Therefore, the Watershed algorithm was selected due to its ability to separate objects based on intensity variations and object contours, even when boundaries are unclear or blurred.

The research process begins with image data collection from a publicly available Kaggle dataset containing various Padang food images. The preprocessing stage includes RGB to grayscale conversion, Gaussian blur filtering, and histogram equalization to enhance image quality and reduce noise. Subsequently, thresholding is applied to produce binary images, followed by distance transform to identify object cores. Marker determination is then performed to distinguish foreground and background regions, which serve as the basis for the Watershed segmentation process.

The Watershed algorithm operates by simulating water flooding from predefined markers until meeting points form object boundaries. Experimental results show that the method can generate clear separation lines between food objects in visually complex scenes. However, quantitative evaluation reveals that the segmented foreground area remains relatively small, indicating that further optimization is required. Overall, the Watershed method demonstrates potential for handling overlapping objects and unclear boundaries, and can serve as a foundation for future culinary image analysis systems.