Pornographic image recognizing method based on sensitive parts detection

Abstract

The invention discloses a pornographic image recognizing method based on sensitive parts detection, comprising the following steps of: step 1, a stage of training a model: constructing five sets of image libraries, training a skin color predicating model, a mammary areola detector, a pudendum detector, a mammary areola identifier and a pudendum identifier; step 2, a stage of pre-processing and extracting a region to be detected: using the skin color predicating module to pre-process the skin color of the image to be detected, and extracting the region to be extracted; step 3, a stage of detecting the sensitive parts: using a slide window to judge whether some sub-region in the region to be detected belongs to the suspected sensitive parts, and identifying and judging the suspected sensitive parts. By the method of the invention, the characteristics such as colors, textures, edges, expressions and the like are blended to a learning-based image identifying frame so as to realize quick and synchronous detection for the sensitive parts, and the pornographic image recognizing accuracy rate and recognizing efficiency are improved.

Description

technical field [0001] The invention relates to an image recognition method, in particular to a pornographic image recognition method based on detection of sensitive parts of a human body. Background technique [0002] After the skin color detection, the traditional pornographic image detection technology will perform one of the following two detections: (1) model-based detection, which mainly detects the human torso; (2) region-based detection, which extracts information such as shape, features such as textures. The main problem of these methods is that they do not detect sensitive parts of the human body, and large areas of skin color do not represent pornographic images (such as bikini images), which will lead to serious misjudgment (such as the detection method of "Green Dam · Flower Season Escort" will misjudge cartoon images such as Garfield as pornographic images). [0003] In the field of object detection, methods based on statistical learning have been shown to ha...

AI Technical Summary

Problems solved by technology

However, these two methods have the following problems: (1) the detection rate needs to be improved

(2) Low detection efficiency

Since the Viola-Jones framework uses an exhaustive search strategy for searching, the detection speed decreases rapidly as the image size increases, making these methods difficult to be practical

(3) The method of eliminating false positives is not reliable enough

The areolas in the actual pictures are very different. Once the above artificially defined priors are not satisfied, the algorithm will fail

(4) There are many empirical parameters

The former sets three ratio thresholds; the number of Gaussian density functions in the skin color model used by the latter is an empirical value, which is difficult to adapt to changes in skin color distribution caused by different lighting conditions and different camera parameters; The skin color detection ratio threshold of is also an empirical value. In fact, this ratio in many pornographic images may be smaller than this value, so a large number of pornographic images will be missed.

(5) The vagina was not tested

Existing methods cannot filter out a pornographic image with only pussy and no areola

Images