Existing Technologies 现有技术

Skin-Tone Based Methods 基于肤色的原理

Most of porn-detection or porn-removal software use color skin tone to detect the spatial distribution of skin area and guess the possibility of the nudity of the scanned images. One example is Snitch's SkinScan algorithm that detects adult images by assessment of skin-tone levels and distribution. The drawbacks of this kind of method are: 绝大多数的色情内容侦测/色情内容剔除软件运用了肤色的原理来探测身体不同部位、区域的肤色，根据肤色的不同来猜测扫描图像裸体的可能性。Snitch’s SkinScan算法是一个例子，它通过评估肤色的不同色浓度和不同的分布位置来侦测成人图像。这种方法的缺点是：

•  Very large false-alarm rate. Since in natural, there are many objects or surfaces are of the skin color tone and the diversity of human skin color tones must introduce very big false-alarm rate for this kind of software. As an example, let us take a look at the searching results of the “Snitch Professional”. The search results are displayed in an order sorted by SkinScan; namely, the likeliness generated by the skin color tone. 很高的错报率：由于在自然界中，有很多物体和表面的颜色和肤色接近，并且，不同人种的肤色也有很大的差异，这样的话，这种软件势必会引起很高的错报率。举一个例子，让我们看一看“Snitch Professional”的搜索结果，搜索结果是按照SkinScan的分类顺序来显示的，也就是说，是根据肤色的色浓度不同的可能性产生的结果。
The following listed those images detected by Snitch to be the most-likely pornographic images. The real porn images are blocked in white while the false alarms are shown in whole. Observe that some of the alarms are entirely surprises. Can you imagine the connection between a bowl of food as shown in the last image in the second row and any illicit images? Well, this is a typical failure of skin-tone based porn-detecting algorithms because many things around us have skin-tone colors. 下面的图像是由Snitch侦测出来的最有可能的色情图片，真正的图像色情部位由白色区域阻截掉了，而报警却在整体部分。观察其中的一些报警实在让人惊讶，你能想象一碗食物（位于第二行的最后一个图像）和违禁图片之间有什么联系吗？这就是基于肤色不同原理来进行色情内容侦测算法的一个典型的失败案例，因为我们周围的很多东西具有皮肤的颜色。



Some other surprises were shown in the following screenshot. The same reason behind the false detection of a tank in the field as shown in the last image in the first row. From the point of view of image understanding technologies, this kind of detecting method can only achieve very low automation because a human operator much be in present to do the final judgments between innocent and suspect images. 下面罗列了一些其他的令人惊讶的结果，田野中的坦克战车（第一排最后一个图片）也是同样的原因被错误报警。从图像理解技术的观点来看，这种侦测技术只能在很低程度上做到自动化，因为在这种情况下，必须要有一个操作者在现场做最后的判定来区分无辜的和可疑的图片。



•  Blind to pornography in black and white.. Who told us that pornographies are all in color? In face, many of them are in black and white (or in gray-scale if you prefer jargons from digital image processing). Obviously, all skin-tone based porn-detecting software packages fail to detect in such cases. However, the image-understanding based software such as PNWatch can still function well. 对于黑白的色情图片无法辨识：谁告诉我们所有的色情图像都是彩色的？对于脸部来说，绝大多数都是黑白的（或者是灰色的，如果你选择经过数据图像专业处理的话）。很明显地，所有基于肤色不同原理来进行侦测色情内容的软件包在这种情况下都不能进行有效侦测。然而，基于图像理解的侦测软件还可以很好的工作，比如：软件PornSeer.
A third party review of another typical flesh tone based porno image detection software can be found at 一个第三方也评估了另外一种典型的基于肉体色度的色情图像探测软件，点击 [here] 可以获得该评估， Just in case you can not visit it, the cached version is at 如果你不能访问该链接，可以点击 [here](cached on 04/19/2006).

The following screenshots are taken from PixAlert Auditor 3.1. Note that this software is also most likely based on skin-tone methods. Observe from the detected results that many innocent images are misclassified only because they had many skin color regions. 下面的屏幕截图是从PixAlert Auditor 3.1.取下的，请注意，该软件也极有可能运用了基于肤色不同的原理方法。观察侦测结果，许多无辜的图片被错误地分类仅仅是因为它们有许多与肤色接近的区域。



Look at the rifle in the first image in the second row, what a surprise! Yes, this kind of software just doesn’t understand the image it is looking at. 请看第二排第一幅图片的步枪，这是多么让人奇怪的结果！是的，这种软件根本就不理解图片的内容。

None Skin-Tone Based Methods 非基于肤色色度来侦测的方法

Skeptic's image composition analysis (ICA) is a method based solely on the pattern recognition ability of neural network to detect pornographic images. The features it uses are shapes, textures, and etc. It was report that this technology can distinguish pose, facial expression, clothing and position of bodies. Still, as all the artificial neural network-based methods, the trained artificial neural networks will never reveal the inner knowledge structures of the images to the users. Therefore, large false alarm rate should be experienced. Skeptic's图像综合分析(ICA)采用的方法是基于神经网络的认知能力来侦测色情图片。它所使用的特点是：外形，纹理，等等。据报告，这种技术可以区分姿势、面部表情、衣服和身体的位置。尽管如此，由于是基于人工的神经网络方法，经过训练的人工神经网络永远不可能对用户揭示图像的内部知识结构。因此，在这种情况下也会有很高的出错率。

Image Understanding Based Methods 基于图像理解的方法

Concerning porn-detection tasks, the difference between a machine and a person is that the person knows that only some areas of skin-color tone don’t necessary mean a pornographic scene. Many other attributes such as exposure breasts and penises contributing much more to pornographic scenes. A people understand what kind of attributes to look for in pornographic scenes. This is called “image understanding” in the jargons of digital image processing. Well, on the other hand, since “image understanding” is ill-defined in the principle of digital image processing, the technology behind image-understanding based porn-detection is not yet another “simple” image understanding methods that spread around us for a long time, it is called Physical Linguistic Vision Technologies. The software PNWatch uses much more advanced image understanding technologies such as breast-detectors and penis-detectors to feature the essential elements that constitute a pornographic image. For more information on how PNWatch works, click 提到色情内容的侦测任务，机器与人之间的区别在于：人能知道在某些具有肤色的区域没有必要视作是色情场景， 还有其它许多的地方如暴露的胸部和阴茎则会很大程度上归类为色情内容。人能够理解、 分辨在什么样的情况下判定为色情场景。这就是所谓的在数字图像处理中的术语 “图像理解”。 另外一方面，由于在数字图像处理原理中对“图像理解”的错误定义， 在基于图像理解原理上的色情内容侦测软件背后的技术不是另外一个在我们周围流传很久的 那个图像理解方法，它叫做物理语义视觉技术。PNWatch软件运用了许多非常先进的图像理解技术， 比如胸部探测器和阴茎探测器用来作为构成一幅黄色图片的基本元素特征。关于PNWatch如何工作的更多信息，请点击 here.