基于感兴趣区域检测的网络不良图片识别研究
发布时间：2018-08-01 13:59
【摘要】：互联网中色情图片传播泛滥,对其自动识别与过滤越来越重要。在本课题中,主要针对网络上常见的单人色情写真类图片,提出了基于感兴趣区域(Regions of Interest,ROIs)检测的不良图片识别算法。传统的不良图像检测算法主要将人体皮肤部分作为感兴趣区域,从皮肤检测的结果中提取与肤色相关的一些信息,如肤色像素所占面积比例等,再结合皮肤的颜色、纹理、形状等特征进行分类。这种方法虽然简单直观,但是在保证较高正检率的前提下,往往误检率也往往较高,尤其对于类如泳装模特等裸露较多的正常图片,效果不甚理想。我们在总结了已有方法不足的基础上,提出了将人体躯干部位作为感兴趣区域的不良图片检测方法。首先使用基于Poselet姿态部件的人体躯干检测方法定位出与色情信息密切相关的躯干区域,结合此兴趣区域和SIFT特征训练高斯混合模型,获取具有判别力的Fisher向量,再利用SVM学习算法训练得到裸露胸部的分类器。然而,由于人体外观变化很大,躯干检测器输出的置信度最大的位置往往较躯干真实的位置有一定的偏移。为了克服这一缺点,我们进一步提出了一种自适应的算法,即根据躯干检测器输出的置信度自适应的选择多个躯干候选区域,并通过集成多个区域的判别结果来得到最终结果。此外,为了训练基于躯干的SVM分类器和验证算法的有效性,本文通过互联网下载的方式收集了一个包含30,000幅单人色情写真图片的大规模数据集,并对色情部位进行了标注,标注信息可用于自动生成训练数据。本文提出的基于躯干的自适应分类算法在收集的大规模数据集上达到了91.7%的识别精度,明显高于传统肤色模型的识别结果。文中采用的基于姿态部件的感兴趣区域检测能够获取与色情信息更相关的信息,因而相比较于传统方法,在较为准确地检测不良图片的同时,有效地降低皮肤裸露较多的正常图像的误检率,达到了实际应用的要求。
[Abstract]:The spread of pornographic images in the Internet is becoming more and more important for its automatic recognition and filtering. In this paper, aiming at the common single person pornographic pictures on the network, a bad image recognition algorithm based on (Regions of Interestrois detection is proposed. The traditional bad image detection algorithm mainly takes the human skin as the region of interest, extracts some information related to skin color from the results of skin detection, such as the proportion of skin color pixels to the area, and then combines the skin color, texture, etc. Shape and other features are classified. Although this method is simple and intuitive, under the premise of ensuring higher positive detection rate, the false detection rate is often higher, especially for the more exposed normal pictures such as swimsuit models, the effect is not very good. On the basis of summarizing the shortcomings of the existing methods, we propose a method for detecting the bad images of the region of interest by taking the trunk part of the human body as the region of interest. Firstly, the human torso detection method based on Poselet pose component is used to locate the torso region which is closely related to pornographic information. Combined with this region of interest and SIFT features, the mixed Gao Si model is trained to obtain the discriminant Fisher vector. Then the SVM learning algorithm is used to train the bare chest classifier. However, due to the great changes in human appearance, the position of maximum confidence in the output of the trunk detector is often offset to the true position of the trunk. In order to overcome this shortcoming, we further propose an adaptive algorithm, which adaptively selects multiple trunk candidate regions according to the confidence output of the trunk detector, and obtains the final results by integrating the discriminant results of multiple regions. In addition, in order to train the trunk based SVM classifier and verify the validity of the algorithm, this paper collects a large scale data set including 30000 portraits of a single person by the way of Internet download, and marks the pornographic parts. Annotated information can be used to automatically generate training data. The self-adaptive classification algorithm based on trunk in this paper achieves 91.7% recognition accuracy on the collected large-scale data set, which is obviously higher than the recognition result of traditional skin color model. In this paper, the region of interest detection based on attitude components can obtain more relevant information than traditional methods, so it is more accurate to detect bad images at the same time. It can effectively reduce the false detection rate of normal images with more skin exposure and meet the requirements of practical application.
【学位授予单位】：南京航空航天大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
，
本文编号：2157803
[Abstract]:The spread of pornographic images in the Internet is becoming more and more important for its automatic recognition and filtering. In this paper, aiming at the common single person pornographic pictures on the network, a bad image recognition algorithm based on (Regions of Interestrois detection is proposed. The traditional bad image detection algorithm mainly takes the human skin as the region of interest, extracts some information related to skin color from the results of skin detection, such as the proportion of skin color pixels to the area, and then combines the skin color, texture, etc. Shape and other features are classified. Although this method is simple and intuitive, under the premise of ensuring higher positive detection rate, the false detection rate is often higher, especially for the more exposed normal pictures such as swimsuit models, the effect is not very good. On the basis of summarizing the shortcomings of the existing methods, we propose a method for detecting the bad images of the region of interest by taking the trunk part of the human body as the region of interest. Firstly, the human torso detection method based on Poselet pose component is used to locate the torso region which is closely related to pornographic information. Combined with this region of interest and SIFT features, the mixed Gao Si model is trained to obtain the discriminant Fisher vector. Then the SVM learning algorithm is used to train the bare chest classifier. However, due to the great changes in human appearance, the position of maximum confidence in the output of the trunk detector is often offset to the true position of the trunk. In order to overcome this shortcoming, we further propose an adaptive algorithm, which adaptively selects multiple trunk candidate regions according to the confidence output of the trunk detector, and obtains the final results by integrating the discriminant results of multiple regions. In addition, in order to train the trunk based SVM classifier and verify the validity of the algorithm, this paper collects a large scale data set including 30000 portraits of a single person by the way of Internet download, and marks the pornographic parts. Annotated information can be used to automatically generate training data. The self-adaptive classification algorithm based on trunk in this paper achieves 91.7% recognition accuracy on the collected large-scale data set, which is obviously higher than the recognition result of traditional skin color model. In this paper, the region of interest detection based on attitude components can obtain more relevant information than traditional methods, so it is more accurate to detect bad images at the same time. It can effectively reduce the false detection rate of normal images with more skin exposure and meet the requirements of practical application.
【学位授予单位】：南京航空航天大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
，
本文编号：2157803
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2157803.html
