基于仿生视觉的图像RST不变属性特征提取方法  

RST-invariant feature extraction method inspired by bionic visual perception

作　　者：余伶俐[1] 夏旭梅 周开军[2] 陈海初[3] 

机构地区：[1]中南大学信息科学与工程学院,长沙410083 [2]湖南商学院计算机与信息工程学院,长沙410205 [3]南昌大学机电工程学院,南昌330031

出　　处：《仪器仪表学报》2017年第4期985-995,共11页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(61403426;61304253);国家重点实验室开放基金(SKLMT-KFKT-201602;SKLRS-2017-KF-13)项目资助

摘　　要：针对图像目标识别过程中易受旋转、缩放、平移及噪声影响的问题,提出一种仿生物视觉感知的RST不变属性特征提取方法,以提升形变目标的识别率与抗噪鲁棒性。受生物视觉感知机理启发,其皮质细胞经过多级变换后,能够最佳权衡图像选择性与不变性。为此,该方法设计成两个阶段。第1阶段中,受生物视觉在水平与垂直方向响应强烈的启发,提出Gabor滤波器与双极滤波器融合的filter-to-filter方向边缘检测方法。Gabor滤波作为底层滤波器平滑图像,通过高层水平与垂直方向双极滤波器检测边缘,构建方向边缘检测子。以增强特征提取的鲁棒性,提升边缘检测的准确度。在此基础上,模拟大脑视觉皮质细胞对线条响应强度的反馈,根据不同边缘方向及间距,度量图像线条的空间频率。设计空间频率间距检测子,将方向边缘图像映射至方向θ-间距I坐标系中。使原图像的旋转与比例缩放,在该坐标系上表现为水平与垂直方向变化。在第2阶段中,针对第1阶段输出图像,再次进行方向边缘检测与间距检测。将第1阶段中水平与垂直平移变换,转变为第2阶段的特征图中不变像素点,使图像具有RST不变性。通过实验统计分析,验证了本文特征的RST不变性及其识别能力。并与其他不变属性特征提取方法进行了识别率与复杂度比较,突显本方法对噪声的强鲁棒性与RST的高识别率。Rotation,scaling and translation（ RST） invariant features extraction with bionic vision mechanism is proposed to improve their recognition accuracy and robustness. Inspired by the biological visual perception,the cortical cells are able to balance the selectivity and invariance of the image with multiple transformations. Therefore,the proposed method is devided into two stages. In the first stage,inspired by the horizontal and vertical directions response of biological vision,a novel filter-to-filter orientation edge detector is built that combines Gabor filters and bipolar filters. The Gabor filters are used as the bottom filter to smooth images,and the edge detector is constructed by the edge of the horizontal and vertical bipolar filter,to enhance the feature extraction robustness and the edge detection accuracy. On this basis,response intensity of cortical visual cortex cells are simulated and the spatial frequency of image is measured according to the different edge direction and distance. Furthermore,spatial frequency interval detector is designed by orientation-interval image mapping inθ-I coordination,which transforms rotation and scaling of original image into a horizontal or vertical shift. In the second stage,the orientation and interval detection are performed once again on the output of first stage,which converts the horizontal and vertical shift into an invariant pixel in orientation-interval map to make the imaoge RST invariance. Experimental results illustrate the effectiveness of invariant features and recognition ability. Meanwhile,the recognition accuracy and complexity are compared with other invariant feature extraction methods,which shows the proposed method is superior on robustness to rotation,scaling,translation and noise image.

关 键 词：RST不变属性特征 仿生物视觉感知 图像特征提取 目标识别 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]