基于多分类和ResNet的不良图片识别框架

针对实际应用中色情图片的复杂多样性问题，提出一种基于多分类和深度残差网络（ResNet）的不良图片识别框架.不同于已有的方法将色情图片识别作为二分类问题，该方法基于多样性特征将色情图片分为7个更细粒度的类别，并将正常图片分为是否包含人物2个类别，通过50层ResNet模型进行分类，再按照阈值计算是否属于不良图片.为了减少训练时间和挖掘优质特征，采用一种反馈修正的训练策略.提出一种单边滑动窗口的预处理方法以解决图片不同尺度的影响问题.测试结果表明，该方法在时间效率和识别准确率上效果良好.

Pornographic Images Recognition Framework Based on Multi-Classification and ResNet

To filter the variety of pornographic images in the reality Internet, the study proposed a Pornographic Images Recognition (PIR) framework based on multi-classification and deep Residual Network (ResNet). Traditional methods usually consider the PIR task as a binary classification, while the approach presented in this paper divides porno images into 7 detailed classes based on its variety features with 2 more benign image classes (with or without human in it). The approach relies on 50-ResNet to extract image features automatically, and then decides whether it belongs to porno images based on the highest score and gives threshold value. At training stage, a feedback-reconstruct training tactics is adopted for the network to collect better features. To deal with images in different scales, a monolateral sliding window method is taken to get better performance. After testing on the data set constructed with collected images from the Internet, the experimental result shows that the approach can reach high accuracy with lower time cost.