CAJ | 학술논문

针对现有特征提取方法难以实现从含有复杂背景的图像中提取有用目标特征的瓶颈问题，提出了基于弱监督学习的去噪受限玻尔兹曼机特征提取算法。首先，利用训练样本，通过无监督学习方式训练一个标准受限玻尔兹曼机模型，从而获得一个包含可视单元层和隐藏单元层的层次结构模型；然后，对可视层的每个单元引入二值转换单元，对隐藏层，根据各节点的激活值大小和激活频率将其分为两组：前景特征隐层单元和背景特征隐层单元，得到一个二元混合式去噪玻尔兹曼机的模型；最后，通过多模交互方式，利用有限数量的样本标签信息对输入样本逐像素地进行采样训练，以此来提取目标特征。实验表明，本文的特征提取算法能够有效地从复杂的干扰背景中提取目标特征，提高了目标识别精度。
침대현유특정제취방법난이실현종함유복잡배경적도상중제취유용목표특정적병경문제，제출료기우약감독학습적거조수한파이자만궤특정제취산법。수선，이용훈련양본，통과무감독학습방식훈련일개표준수한파이자만궤모형，종이획득일개포함가시단원층화은장단원층적층차결구모형；연후，대가시층적매개단원인입이치전환단원，대은장층，근거각절점적격활치대소화격활빈솔장기분위량조：전경특정은층단원화배경특정은층단원，득도일개이원혼합식거조파이자만궤적모형；최후，통과다모교호방식，이용유한수량적양본표첨신식대수입양본축상소지진행채양훈련，이차래제취목표특정。실험표명，본문적특정제취산법능구유효지종복잡적간우배경중제취목표특정，제고료목표식별정도。
Existing feature extraction algorithms are difficult to capture useful information from complex images .A feature extraction approach is proposed based on the weakly supervised learning with denoising restricted Boltzmann machine(RBM).First, a standard RBM is pre-trained in an unsupervised learning way,which provides a hierarchical mode with a visible layer and a hidden layer.Second,for the visible layer,a stochastic binary switch node is employed.And for the hidden layer,it is divided into fore-ground-hidden nodes and background-hidden nodes based on the score of each hidden node’s activation values and times,thus we can achieve a binary mixture denoising RBMs .Finally,the pixel-wise denoising RBMs is trained by using small number label infor-mation and stochastic switch nodes through multiplicative interaction .The experimental results show that significant performance im-provement is achieved with our proposed method .