基于幼苗生长点与盘穴中心匹配的番茄苗盘缺苗检测方法

Missing Seedling Detection Method for Tomato Plug Trays Based on Matching of Seedling Growth Points and Cell Centers

摘要:

针对设施番茄育苗过程中,幼苗（7~14 d）因苗茎弯斜造成俯视角度下苗叶侵入邻近盘穴,难以通过垂直投影方法准确进行缺苗检测的问题,本文提出了基于幼苗生长点与盘穴中心匹配的缺苗检测方法。通过引入轻量化C3K模块、注意力机制和替换损失函数的方法对YOLO 11改进,提出TSM-YOLO模型以提升对番茄幼苗的实例分割性能;基于实例分割所得的掩码,采用质心法计算其形态学中心,并通过坐标变换得到幼苗生长点的实际坐标,并基于直通滤波和直方图方法对盘穴中心进行分析计算,实现幼苗生长点和盘穴中心点同步精准定位;通过改进匈牙利算法实现2类中心点匹配,未成功匹配的盘穴中心点即为缺苗位置。试验结果表明,TSM-YOLO模型实例分割指标AP50和AP50-95分别达到97.6%和78.9%,对比基础模型分别提高2.4、11.4个百分点,模型参数量、浮点运算量和内存占用量分别减少7.1%、8.1%和8.8%,推理速度提升12.1%;幼苗生长点定位试验对比多种算法,其中质心算法定位偏差最大值为8.09 mm,平均值为3.39 mm,标准差为1.78 mm,均为最优,更适用于生长点定位工作;优化后的匈牙利算法子叶期匹配正确率为99.9%,单盘用时8.3 ms,出苗早期匹配正确率为90.7%,单盘用时8.5 ms,相比传统匈牙利算法稳定性更强,效率更高。通过设置对比试验比较目前主流视觉检测方法与本文方法对不同生长时期番茄苗盘的检测情况,结果表明本文方法在子叶期检测正确率达98.1%,漏检率1.9%,误检率0.3%,出苗早期成功率达91.4%,漏检率8.6%,误检率1.0%,具备更强的复杂环境适应性和检测正确率。

Abstract:

Aiming to address the issue of inaccurate seedling vacancy detection in greenhouse tomato propagation where 7~14-day-old seedlings often lean or bend, causing leaves to intrude into adjacent tray cells and rendering traditional vertical projection methods ineffective, a detection method based on the matching of seedling growth points with tray cell centers was proposed. Firstly, the TSM-YOLO model was developed by improving YOLO 11 through the introduction of lightweight C3K modules, attention mechanisms, and optimized loss functions to enhance instance segmentation performance for tomato seedlings. Secondly, based on the masks obtained from instance segmentation, the centroid method was employed to calculate morphological centers, which were transformed into actual growth point coordinates;meanwhile, tray cell centers were precisely located through straight-through filtering and histogram-based analysis to achieve synchronous positioning of both center types. Finally, an improved Hungarian algorithm was implemented to match the growth points with cell centers, where unmatched cell centers were identified as vacant positions. Experimental results showed that the TSM-YOLO model achieved an AP50 and AP50-95 of 97.6% and 78.9% respectively, representing increases of 2.4 and 11.4 percentage points over the base model, while reducing parameters by 7.1%, FLOPs by 8.1%, and model size by 8.8%, with a 12.1% boost in inference speed. Seedling growth point localization experiments were performed to compare different algorithms. The centroid algorithm obtained the best performance: maximum localization deviation of 8.09 mm, mean of 3.39 mm, standard deviation of 1.78 mm. For the optimized Hungarian algorithm, matching accuracy was 99.9% (8.3 ms per tray, cotyledon stage) and 90.7% (8.5 ms per tray, early emergence stage), demonstrating higher stability and efficiency than the traditional version. Comparison with mainstream vision detection methods indicated that the proposed method reached a detection accuracy of 98.1% with a missed detection rate of 1.9% and a 0.3% false detection rate at the cotyledon stage, and 91.4% accuracy with a missed detection rate of 8.6% and a 1.0% false detection rate at the early emergence stage, exhibiting superior adaptability to complex environments and higher overall detection precision.