级联型P-RBM神经网络的人脸检测

Cascaded probability state-restricted Boltzmann machine for face detection

目的针对非理想条件下快速准确的人脸检测问题,提出一种基于概率态多层受限玻尔兹曼机(RBM)级联神经网络的检测方法。方法它采用RBM中神经元的概率态表征来模拟人脑神经元连续分布的激活状态,并且利用多层P-RBM(概率态RBM)级联来仿真人脑对视觉的层次学习模式,又以逐层递减隐藏层神经元数来控制网络规模,最后采用分层训练和整体优化的机制来缓解鲁棒性和准确性的矛盾。结果在LFW、FERET、PKUSVD-B以及CAS-PEAL数据集上的测试都实现了优于现有典型算法的检测性能。对于单人脸检测,相比于Adaboost算法,将漏检率降低了2.92%;对于多人脸检测,相比于结合肤色的Adaboost算法,将误检率降低了14.9%,同时漏检率降低了5.0%,检测时间降低了50%。结论无论是静态单张人脸,还是复杂条件下视频多人脸检测,该方法不仅在误检率和漏检率上表现更好,而且具有较快的检测速度,同时对于旋转人脸检测具有较强的鲁棒性。针对基于肤色的多人脸检测研究,该方法能显著降低误检率。

Objective Face detection is constantly an active research subject in computer vision and pattern recognition. Face detection is also a constituent part of pattern recognition, artificial intelligence, information security, and many other disciplines. With video network coverage widely increasing in recent years, face detection has been increasingly used in the field of video surveillance. However, many factors require consideration in face detection, such as the complex environ- ments, multiple faces, and face rotation angles. In view of these interference problems in nonideal condition, a cascaded neuron network based on a multi-layer probability state-restricted Bohzmann machine （P-RBM） is proposed in this study to overcome the challenge of accurately and rapidly detecting faces. Method The neurons of RBM only have two states, name- ly, activated and nonactivated; this state mode can inhibit the interference in the learning result induced by the inadequate active information, while it simultaneously increases the likelihood that the learning network falls into a local optimum caused by the shielding of relatively weak information. To solve this contradiction, the proposed method uses the probability state of neurons in RBM as their activation degree, which better models the activity state＇ s continuous distribution of the neurons in the human brain. Using the probability state not only retains the weak active information but further decreases the effect caused by the former layer＇ s miscalculation. Simultaneously, this method simulates the hierarchical learning mode in the human brain by cascading multiple P-RBMs. This cascaded network can achieve multi-layer nonlinear mapping and obtain the semantic feature of the input date by extracting the input data＇ s separate level features. Furthermore, this cascaded network can learn the relationship hiding within the data to make the learned features be more promotional and ex- pressive. Simultaneously, the number of the hidden layer＇ s neurons decreases layer-by-layer to control the network＇ s scale and enhance the robustness. Finally, the proposed method uses the layered training and the entire optimization to balance robustness and accuracy. The greedy layer-wise learning is used in the layered training to avoid the training error transfer- ring in layers, thereby solving the problem of the multi-layer network easily falling into the local optimum. Furthermore, a preprocessing layer is used to detect the skin color area to reduce the number of neurons in the detection network and speed up the detection speed. Result Testing the single face detection performance in the LFW and FERET, the proposed method nearly achieves entirely accurate detection. Testing the video face detection in the PKU-SVD-B database, the missing de- tection rate and the false detection rate of the proposed method are all lower than that of the state-of-the-art methods, such as Adaboost and Adaboost combined with skin color detection, and its detection speed is faster. Moreover, the proposed method has a good detection performance for the face with a large rotation, which is tested in the CAS-PEAL database. Conclusion Experimental results show that regardless of whether a static single face or video multi-face detection occurs under complicated conditions, apart from the faster detection speed and robustness against face rotation, the proposed method possesses lower false detection rate and lower missing detection rate. Aiming at the multi-face detection based on skin color, this method can significantly reduce the false detection rate.