基于机器视觉的温室番茄裂果检测

该文通过对温室番茄果实进行定位及裂果检测,可为番茄裂果率预估及后续裂果自动筛选提供参考。针对自然光照下采集的各类番茄图像,在相关颜色空间中进行阈值预分割,利用前期支持向量机训练得到的纹理特征分类器对预分割区域进行二次判别;之后在前景区域利用显著性角点分割构造边缘轮廓集,利用基于最小二乘法修正的改进霍夫变换拟合单个番茄目标;最后利用二维Gabor小波算子对拟合的单个番茄区域进行纹理特征提取及裂果判别。文中共采集82幅番茄图像,其中50幅图像作为训练集图像,32幅图像作为验证集,所提算法对测试集中总共128个番茄的果实正确检出率为91.41%,对其中35裂果的正确判别率为97.14%,裂果判别部分平均耗时21 ms。试验结果表明,该方法具有较好的鲁棒性与可靠性,对成熟期番茄裂果率的估计研究及采摘过程中裂果的自动分级筛选具有较好的指导意义,为未来实现温室番茄果实生长状态在线监测提供参考。

Detection of cracking tomato based on machine vision in greenhouse

Abstract: A new combined algorithm is put forward to facilitate the prediction of tomato cracking rate and automatic screening of dehiscent fruit. In order to improve the recognition accuracy and reduce the segmentation error in natural illumination, different color spaces of the original image were compared in the preliminary segmentation section, then multi-channel in color space that including R-Bchromatic aberration characteristic, normalized R channel and Hue channel were chosen. For the pre-segmentation may include some non-target areas, the relevant texture features were used to make a secondary identification of potential areas. In this study, SVM (support vector machine) was built based on fruit areas and non-fruit areas of a certain size (10×10 pixels) extracted from the training image. 5 texture features, including standard deviation, smoothness, third-moment, energy, and entropy were calculated for those fruit areas and non-fruit areas, thus the regions of target and background could be successfully separated by the algorithm. Then, the edges and contours, extracted in this foreground area, were used to construct the contour dataset. The Shi-Tomasi corner detection algorithm was implemented to split the contours in this dataset. Since the edges of the tomato fruit were mainly arc fragments, the contour set was preliminary selected according to the contour length and contour curvature. This part was especially important to simplify the contour set and improve the efficiency of subsequent calculation. Circular Hough transform (CHT) was then applied to fit the contour set. The maximum value of distance transform in foreground binary region was taken as the limit of fitting ellipse radius. If the circle radius was bigger than the maximum fruit radius, the circle would be rejected. If the distance between 2 circles was smaller than two-thirds of the maximum value, the circle would be rejected due to the heavy occlusion between 2 tomato fruits. The least square contour correction was made based on the roundness and the number of background pixels contained in this ellipse area. The best results were thus selected from multiple fitting results in the same region. The proposed method combined the texture, color, and shape information of the tomatoes and presented a good recognition accuracy in greenhouse. In view of the great difference in texture features between good fruit and dehiscent fruit, texture feature was selected in this study. Two-dimensional Gabor wavelets transform can extract texture feature from different scales and different directions, which is also insensitive to illumination and rotation. Therefore, the Gabor wavelets was used to distinguish good fruit and dehiscent fruit. Texture features including the energy and normalized mean were extracted from 4 scales and 10 directions in good fruit and dehiscent fruit regions in the training images, which contained about 195 good fruit regions and 55 dehiscent fruit regions. Then another SVM classifier was trained based on this texture feature to distinguish the recognized fruits in subsequent experiment. A total of 82 images were used in this study, in which 50 images were used as training images, and the other 32 images were used as validation images. Experiments showed the correct recognition rate for 128 tomato fruits in the total 32 images is 91.41%, the recognition rate for the dehiscent fruits reached 97.14%, the average processing time of this algorithm was 249 ms. This algorithm had good robustness, stability for fruit recognition and dehiscent fruit identification, which was instructive for the estimation of tomato yield and automatic classification of dehiscent fruit in the process of picking system. It would also build a solid foundation for the future implementation of on-line monitoring system in greenhouse, whichwas used to record growth information during the plant growth cycle.