基于仿生视觉的图像RST不变属性特征提取方法

发布时间：2018-05-20 22:01

本文选题：RST不变属性特征 + 仿生物视觉感知　；参考：《仪器仪表学报》2017年04期

【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。
[Abstract]:In order to improve the recognition rate and robustness of deformed targets, a RST invariant feature extraction method based on biologic vision perception is proposed to solve the problem that image target recognition is easily affected by rotation, scaling, translation and noise. Inspired by the mechanism of biological visual perception, the cortical cells can best balance image selectivity and invariance after multilevel transformation. For this reason, the method is designed in two stages. In the first stage, inspired by the strong response of biological vision in horizontal and vertical direction, a method of edge detection in filter-to-filter direction based on the fusion of Gabor filter and bipolar filter is proposed. Gabor filter is used as the smoothing image of bottom filter. A directional edge detector is constructed by using a high level horizontal and vertical bipolar filter to detect the edge. In order to enhance the robustness of feature extraction, improve the accuracy of edge detection. On this basis, the feedback of the visual cortical cells to the response intensity of the line was simulated, and the spatial frequency of the lines was measured according to the different edge directions and distances. The spatial frequency spacing detector is designed to map the direction edge image to the direction 胃-spacing I coordinate system. The rotation and scale of the original image are changed in horizontal and vertical directions in this coordinate system. In the second stage, direction edge detection and distance detection are performed again for the first stage output image. The horizontal and vertical translation in the first stage is transformed into the invariant pixels in the feature map of the second stage, which makes the image RST invariant. The RST invariance of the feature and its recognition ability are verified by the experimental statistical analysis. The recognition rate and complexity of this method are compared with other invariant attribute feature extraction methods, which show the strong robustness of this method to noise and the high recognition rate of RST.
【作者单位】：中南大学信息科学与工程学院;湖南商学院计算机与信息工程学院;南昌大学机电工程学院;
【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
【分类号】：TP391.41

【相似文献】