利用骨架特征信息的粘连谷粒图像分割方法

Touching Corn Kernels Based on Skeleton Features Information

针对长宽比大的粘连谷粒图像精确分割问题,提出利用图像骨架粘连点位姿信息的粘连谷粒图像分割方法,相比常规分水岭算法提高了准确性和可靠性。在图像平滑、二值化的基础上提取谷粒的骨架信息,采用SPT算子定位骨架粘连点位姿并进行粘连点延展闭合。针对不同粘连程度的谷粒图像,将本文算法与传统分水岭算法进行了对比测试,结果表明本文算法对于复杂粘连情况具有较强的适应性。当谷粒为长宽比在1.5以内的大豆和玉米时,本文方法和分水岭算法效果类似。然而,当谷粒长宽比大于1.5,分水岭算法出现大量欠分割和过分割现象,识别错误率达到75%,本文算法依然可控制在10%以内。

The segmentation method based on Watershed algorithm is inefficient when the object has a much higher aspect ratio. This paper presents an automatic separation procedure of touching kernel images based on their skeleton features. First the images were preprocessed by the vector median filter to get the smoothing images. Then the improved single-pass thinning algorithm was adopted to abstract binary image skeleton. After getting the pose of possible adhesive points with SPT algorithm, each endpoint of skeleton lines concluded and corrected recur sively according to the direction and distance. The algorithm was tested for different grain types under different touching scenarios and was successful in separating more than 90% of the touching grains when classical watershed methods allow only to segment 25% of the high length to width ratio （ 〉 1.5） grains. The algorithm appears to be robust to separate touching scenarios where the kernels have different length-width ratio.