Comparison of SIFT and SURF Methods for Porn Image Detection

Pornography is a global problem. Various efforts to overcome the spread of pornographic videos have been carried out by many researchers in the world with various methods such as the GLCM method for extracting image features based on skin color, the BossaNova method for extracting features in the images, combining two color spaces namely RGB and YCbCr and using segmentation skin color with YCbCr and KNN. Feature extraction used in the previous ones such as GLCM, YCbCr, RGB was vulnerable to changes in image size, translation, and rotation. In contrast to the SIFT and SURF methods that were used in this study. SIFT and SURF method invariant to changes in translation, size, two-dimensional rotation, illumination, can extract many key points and able to show object decryptors better. The number of datasets used was 7997 pornographic and non-pornographic images taken from Yahoo’s dataset (NSW dataset) while 984 images were used as data testing. The feature extraction used the SIFT and SURF methods to get the keypoint and descriptor values. Descriptors of each image were clustered by K-Means. The center of each cluster was used as a visual dictionary’s vocabularies. Then the image representation was performed using BoVW and finally, the classification was done by the KNN algorithm. For the results, the SURF method had better accuracy and time values compared to the SIFT method. The highest accuracy value of the SURF method was in dictionary size 300 with an accuracy of 0. S226 (S2.26%). While the lowest accuracy value was in dictionary size 2600 with an accuracy of 0.6917 (69.17%). For the time variable, the SURF method was slightly faster than the SIFT method with an average value of respectively 2.49 and 2.52