基于相移算法的苹果果梗/花萼检测方法

由于苹果表面缺陷与果梗/花萼具有相似的灰度特征，通过传统机器视觉方法难以对两者进行有效区分。为避免苹果果梗/花萼对其表面缺陷识别造成干扰，该研究提出了一种基于相移算法的苹果果梗/花萼检测方法。通过搭建条纹投影系统，投影仪投射三步相移条纹至苹果样本，摄像机同步采集经苹果表面调制的条纹图像；通过分析发现果梗/花萼区域的条纹图像凹凸性与正常区域存在明显差异，利用三步相移算法恢复条纹图像的截断相位，结合相位偏移、阈值分割和二维凸包算法便可检测出苹果果梗/花萼。试验结果表明：该方法能够有效地区分果梗/花萼和表面缺陷，识别出不同位置和角度的果梗/花萼，整体准确率可达到99.12%；同时能够满足在线检测需求，平均处理时间约为0.479s。该研究可为苹果外观品质检测提供技术支持。

Apple stem/calyx detection based on phase-shifting algorithm

Abstract: Surface defect is one of the most important indexes for quality inspection of apple fruits. However, it is always difficult to distinguish between the surface defect and the stem/calyx of apples using traditional machine vision from the color information, due to the high similarity of the intensity distribution. This study aims to prevent the disturbance of the apple stem/calyx on the surface defect recognition. An effective apple stem/calyx detection was proposed using three-step phase-shifting algorithm, in terms of the three-dimensional (3D) information instead of color information. A typical fringe projection system was built with a digital projector and an industrial camera for the apple stem/calyx detection. The digital projector was used to sequentially illuminate three phase-shifting fringe patterns onto the apple sample during image acquisition. At the same time, the industrial camera was used to synchronously capture three fringe images that modulated by the apple surface. There were the outstandingly different concavity and convexity of the modulated fringe images within the stem/calyx regions from the normal. The reason was that the stem/calyx regions on the apple surface were usually concave relative to the other normal regions. The experimental analysis showed that the modulated fringe images were bending the left within the stem/calyx regions, while bending the right within the normal regions. Three-step phase-shifting algorithm was then utilized to recover the wrapped phase of the modulated fringe images that used to indicate the bending direction. Several shifted wrapped phases were computed from the original wrapped phase with the assistance of the remainder operation. The binary fringes were then obtained by simply applying threshold segmentation on these shifted wrapped phases. The convex residuals on the right side of these binary fringes were extracted using connected component labeling and two-dimensional convex hull algorithms. As such, the entire region of apple stem/calyx was detected to combine all convex residuals of these binary fringes. The total 684 group of modulated fringe images were captured from the apple samples to detect the stem/calyx of different apples with the different sizes, colors or poses, in order to validate the performance of the improved model. These modulated fringe images were mainly divided into four types: normal apple with stem/calyx, normal apple without stem/calyx, defective apple with stem/calyx, and defective apple without stem/calyx. The experiment results indicated that the proposed detection was correctly identified the apple stem/calyx under different positions and angles, even though the apple samples with the outstanding surface defects. Statistical results showed that the overall recognition rate reached 99.12%, and the recognition rate were all beyond 98.30% for the four types of the modulated fringe images. In addition, the average processing time was only about 0.479 s on MATLAB platform, fully meeting the requirement of online detection. Compared with the traditional machine vision using color information, the wrapped phase relating to 3D information was very insensitive to the surface color, indicating more suitable to distinguish the surface defect and the stem/calyx of apples. Moreover, the three fringe patterns were only required to recover the wrapped phase, compared with traditional structured light that require to reconstruct the 3D information of the apple surface. Consequently, the better performance was achieved in the higher recognition rate, the faster image acquisition speed, and the less duration of image processing. The great potential can be expected as the simple setup, high accuracy, and high speed for the stem/calyx detection of various fruits, such as apples, pears, and peaches. The finding can also provide the technical support for the quality inspection of apple surface.