Image Classification by Skin Detection
One interesting application of skin detection is as part of a larger system for detecting people in photos. A person detector that worked reliably on Web images could be a valuable tool for image search services on the web and in digital libraries, as well as for image categorization.
Our goal is to determine whether or not an input image contains one or more people by aggregating the pixel-wise output of the skin detector. The baseline detection rate for this problem is 52%, which is the percentage of images in our dataset containing people. We computed a simple feature vector from the output of the skin detector and then trained a classifier on these features to determine whether a person is present or not. The features we used are
1. Percentage of pixels detected as skin
2. Average probability of the skin pixels
3. Size in pixels of the largest connected component of skin 4. Number of connected components of skin
5. Percent of colors with no entries in the skin and nonskin histograms . RESULTS:
– 83.2% (correctly classified person images)
– 71.3% (correctly classified non-person images)
– 77.5% (correctly classified images)
These features can all be computed in a single pass over the input image, making the resulting person detector extremely fast. No effort was spent tuning or adjusting the feature set, so it is possible that other choices would yield better performance.
1.2 Adult Image Detection
By taking advantage of the fact that there is a strong correlation between images with large patches of skin and adult images, the skin detector can also be used as the basis for an adult image detector. There is a growing industry aimed at filtering and blocking adult content from Web indexes and browsers. All of these services currently operate by maintaining lists of objectionable URL’s and newsgroups and require constant manual updating. An image-based scheme has the potential advantage of applying equally to all images without the need for updating To detect adult images, we followed the same approach as with person detection. A feature vector based on the output of the skin detector was computed for each training image. The feature vectors included the same five features used for person detection, plus two additional elements corresponding to the height and width of the image. These two were added based on informal observations that adult images are often sized to frame a standing or reclining figure. The adult image detector is essentially looking for images with connected regions of skin of the right size. As a consequence, the most common false positive for our system is a close-up image of a face. Use of a face detector in conjunction with the skin detector could alleviate this particular problem. However, major improvements in performance are likely to require the use of other cues such as text, as well as more detailed analysis of image structures. 
We combined the color-based detector (using a threshold that yielded 85.8% correct detections and 7.5% false positives) with the text-based detector by using an “OR” of the two classifiers, i.e. an image is labelled adult if either classifier labels it adult. The combined detector correctly labels 93.9% of the adult images from crawl A and obtains 8% false positives on the non-adult images from crawl B. Table 1 summarizes these results. One interesting observation is that the color and texture features appear to be complementary, since the combined detector exhibits an increase in both the detection rate and the false alarm rate. This suggests that more sophisticated combination schemes than the “OR” operator could yield even better combined performance. 
Table 1. Comparison of adult image detector using color-based, text-based and combined classifiers on the test data [pic]
1.3 Human Face...
Please join StudyMode to read the full document