Exposed body component-based harmful image detection in ubiquitous sensor data

At present, users are able to obtain a variety of image contents easily via visual sensor networks. Under this circumstance, the detection of adult or harmful images has become a highly important issue. This paper proposes an algorithm for analyzing images in the visual sensor network and robustly detecting the human nipple and navel regions, which can be used for harmful image detection. From a color image, the proposed algorithm first detects skin color areas and locates human facial regions, including the eyes and lips. The algorithm then creates a color-based nipple map for detecting candidate regions of nipples from the extracted skin regions. If a detected candidate nipple region is located within the detected facial region, the candidate nipple region is removed because it has been incorrectly detected. Subsequently, by utilizing geometrical features and the mean color filter of the nipples, the method filters the candidate regions of nipples in two steps and detects the actual nipple regions. Lastly, this method exploits the structural relation between the detected nipple regions and the navel areas to be detected, and uses the edge and saturation images to detect the navel region robustly. Experimental results reveal that the proposed adult image algorithm detects the human nipple and navel regions from various types of images more reliably than existing algorithms. Since the proposed approach uses the human structure information of the nipple and navel regions to filter the candidate navel regions effectively, it removes many incorrectly detected regions, which results in high accuracy. We expect that the proposed nipple and navel region detection algorithm will be successfully employed to detect and block harmful images in various real applications.