自适应分层阈值的简化PCNN红外人体图像分割

Simplified Pulse Coupled Neural Network with Adaptive Multilevel Threshold for Infrared Human Image Segmentation

针对红外人体图像中人体目标亮度分布不均匀且易与背景混叠等引起分割不完整的问题,提出一种自适应分层阈值的简化PCNN(SPCNN)红外人体图像分割方法.该方法摒弃了传统SPCNN模型中的动态阈值指数衰减下降机制,采用神经元点火区域与未点火区域的统计特性构建自适应分层阈值;同时结合神经元同步点火机制并引入最近邻均值聚类规则控制神经元点火,以达到较高的分割精度.在真实红外人体图像集上与几种图像分割方法进行对比的实验结果表明,文中方法能取得较优的分割效果以及较小的分类错误率,且与传统的SPCNN模型相比,文中的SPCNN模型参数的设置更加简化.

Aiming at the problem that the incomplete segmentation occurred due to the intensity overlap between background and human targets in infrared human image, simplified pulse coupled neural networks （SPCNN） with adaptive multilevel threshold is proposed for infrared human image segmentation. The proposed method entirely abandons the mechanism of dynamic threshold decaying in time by the exponent term, and constructs multilevel threshold by using the region statistics of fired region and unfired region. Meanwhile, the clustering rule is introduced by combining the synchronous pulse mechanism to regulate the neurons firing in different object regions, so that it is possible to achieve high segmentation accuracy. Compared with several kinds of segmentation algorithms on real- world infrared images, experimental results show the higher efficiency and the lower misclassification of our method for human targets extracting. Furthermore, our model is superior over the traditional SPCNN in terms of parameters setting.