基于稀疏主成分分析和自适应阈值选择的图像分割算法

Novel Image Segmentation Algorithm via Sparse Principal Component Analysis and Adaptive Threshold Selection

图像分割是机器视觉中的基础问题,基于阈值的图像分割算法依赖于参数调整,但参数调整容易受到局部最小值的影响且需要耗费大量时间,从而降低了分割算法的质量和效率。为了实现图像分割过程中的自适应阈值选择,提出了一种基于稀疏主成分分析和自适应阈值选择的图像分割算法。该算法首先利用稀疏主成分分析感知图像的噪声水平以自适应去噪,其次通过二维直方图感知图像的主干区域内容以自适应获得全局分割阈值,然后通过移动平均法的局部阈值分割算法对图像进行分割,最后将全局阈值分割和局部阈值分割图像结合,从而获得最佳的分割图像结果。在伯克利数据集上的仿真实验结果表明:相比传统的阈值分割算法,该算法在分割边缘的准确性和对噪声的鲁棒性上具有一定的优势,在主客观上均具有较好的分割效果,基于稀疏主成分分析的自适应阈值选择方法提高了图像的分割质量。

Image segmentation is a fundamental problem in machine vision. Image segmentation algorithm based on threshold depends on the parameter adjustment,which is vulnerable to local minimum value and needs a lot of time. It reduces the quality and efficiency of segmentation algorithm. In order to realize the adaptive threshold selection in the process of image segmentation, a novel image segmentation algorithm via adaptive threshold selection and sparse principal component analysis was proposed. According to the content of the image, the algorithm removes the noise with the image noise level obtained by the sparse principal component analysis. The global segmentation threshold is obtained by the main region of the image based on 2D histogram. Then the local segmentation threshold is obtained by local details of image based on moving average method. Finally, the global threshold segmentation and the local threshold segmentation image are combined to obtain the best segmentation results. The simulation and experimental results on Berkeley data set show that the algorithm has an advantage on the accuracy edge of image segmentation and robustness to noise compared to current frontier algorithm. It has better segmentation performance on subjectivity and objectivity, and improves the quality of image segmentation.