基于分层PCA技术的显著性目标检测算法

Saliency Detection Based on Hierarchical PCA Technology

摘要: 针对当前显著性目标检测方法存在的背景噪声较大、准确度较低以及计算量大等问题，提出了一种基于分层PCA技术的显著性目标检测算法。该方法首先将原始RGB图像转换为灰度图像，并通过比特面分层技术将原始灰度图像分为8层，每层图像包含与该层图像特征相匹配的显著目标信息；接着，以原图的色彩结构为参考图像，通过灰度级–彩色变换方法为分层后的灰度图像重新赋值，分层图像不仅体现了原有的结构特征而且有效地保留了原始图像的彩色特征；然后，对分层图像进行PCA分析，得到每层图像在主成分方向上的结构差异特征以及色彩差异特征；之后，将两种特征进行融合，产生精度高、鲁棒性好的分层显著图，进而利用中心先验方法将显著目标放置近似中心处从而突出目标区域，得到更加精确的显著图。最后，对分层显著图进行信息熵判决，并由此得到一幅背景信息最少而显著信息突出的最优显著图。在MSRA、ASD、ESSCD等数据集中随机选取300张图像进行了测试并与ITTI (IT)、GBVS (GB)、SR、LC、HS、CHS等几种经典方法进行了比较，实验结果表明所提出方法能够有效的将显著目标与背景相分离，检测效果更接近人工标定的结果，与对照方法相比在准确率(PRE)、召回率(REC)和F-measure等性能参数方面具有明显优势。同时加快了运算速度，提高了检测精确度。

Abstract: Many traditional methods of saliency detection have the disadvantages of intensive background noise, low accuracy and high calculation cost. Therefore, this paper proposes a novel saliency detection algorithm based on hierarchical PCA method. Firstly, according to prominently distinct detail, the original image is divided into eight layers. Each layer contains correlated target information of the image layer shared the same feature. Then, the method of grayscale image colorizing is used to transplant color characteristics from source image to hierarchical grayscale image, which purpose is to make the layered image not only reflect the pattern characteristics but also retain the original color feature. After that, PCA technology is used to detect layered images to obtain distinct object’s pattern distinctness and color distinctness in the principal component direction. Next, two features are integrated to get the saliency map with high robustness, and to further refine our results, the known priors are incorporated on image organization, which can place the subject of the photo-graph near the center of the image. Finally, entropy calculation is used to determine the optimal image from the layered saliency map; the optimal map has the least background information and the most prominently salient target. 300 pictures of the MSRA, ASD and ESSCD databases are randomly selected to test and compare with several classical methods. The layered PCA technology can effectively separate the significant object from the background. The detection results of the proposed method are closer to the manual calibration, while taking advantages of performance parameter include accuracy rate, recall rate and F-measure value. At the same time, it accelerates the calculation speed and improves the detection accuracy.