基于全卷积神经网络与低秩稀疏分解的显著性检测  

Saliency Detection via Full Convolution Neural Network and Low Rank Sparse Decomposition

作　　者：张芳 王萌[2] 肖志涛[1,2] 吴骏 耿磊[1,2] 童军 王雯[1,2] ZHANG Fang;WANG Meng;XIAO Zhi-Tao;WU Jun;GENG Lei;TONG Jun;WANG Wen(Tianjin Key Laboratory of Optoelectronic Detection Technology and Systems,Tianjin 300387;School of Electron-ics and Information Engineering,Tianjin Polytechnic University,Tianjin 300387)

机构地区：[1]天津市光电检测技术与系统重点实验室,天津300387 [2]天津工业大学电子与信息工程学院,天津300387

出　　处：《自动化学报》2019年第11期2148-2158,共11页Acta Automatica Sinica

基　　金：国家自然科学基金(61601325);中国纺织工业联合会应用基础研究项目(J201509);天津自然科学基金(15JCYBJC16600,17JCQNJC01400)资助~~

摘　　要：为了准确检测复杂背景下的显著区域,提出一种全卷积神经网络与低秩稀疏分解相结合的显著性检测方法,将图像分解为代表背景的低秩矩阵和对应显著区域的稀疏噪声,结合利用全卷积神经网络学习得到的高层语义先验知识,检测图像中的显著区域.首先,对原图像进行超像素聚类,并提取每个超像素的颜色、纹理和边缘特征,据此构成特征矩阵;然后,在MSRA数据库中,基于梯度下降法学习得到特征变换矩阵,利用全卷积神经网络学习得到高层语义先验知识;接着,利用特征变换矩阵和高层语义先验知识矩阵对特征矩阵进行变换;最后,利用鲁棒主成分分析算法对变换后的矩阵进行低秩稀疏分解,并根据分解得到的稀疏噪声计算显著图.在公开数据集上进行实验验证,并与当前流行的方法进行对比,实验结果表明,本文方法能够准确地检测感兴趣区域,是一种有效的自然图像目标检测与分割的预处理方法.A unified saliency detection approach via the full convolution neural network(FCNN) and the low rank sparse decomposition is proposed to accurately detect the salient region in complex backgrounds. An image can be decomposed into a low rank matrix and sparse noise, indicating background and salient region, respectively. The highlevel semantic prior knowledge learned by using the full convolution neural network is combined to detect the salient region in the image. Firstly, the original image is clustered into super pixels, and the feature matrix is constructed by extracting color, texture and edge features of each super pixel. Then, the feature transformation matrix is learned with the gradient descent method and the high-level semantic prior knowledge is learned with the full convolution neural network by using the MSRA database. Furthermore, the feature matrix is transformed using the feature transformation matrix and the high-level semantic prior knowledge matrix. Finally, the transformed feature matrix is decomposed into a low rank matrix and a sparse matrix by the robust principal component analysis method, and the saliency map is calculated according to the sparse matrix. The proposed method is compared with state-of-the-art algorithms on the open datasets.Experimental results demonstrate that the proposed algorithm can accurately detect the region of interest, which is an effective preprocessing means for object detection and segmentation of natural images.

关 键 词：显著性检测 全卷积神经网络 低秩稀疏分解 高层语义先验知识 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]