基于改进UNet的复杂环境橡胶树割痕检测与割胶轨线识别方法

Tapping Cut Detection and Tapping Trajectory Recognition of Rubber Tree in Complex Environments Based on Improved UNet

摘要:

天然橡胶树割痕检测与割胶轨线识别对实现割胶自动化、提升作业智能化具有重要意义。针对复杂环境下割胶机器识别割痕、割胶轨线识别易受到自然环境的干扰，导致割痕识别不完整、割胶轨线拟合不精准的问题，提出了一种基于改进 UNet 的复杂环境橡胶树割痕检测与割胶轨线识别方法。针对 UNet 模型识别效率不高的问题，将原 UNet 中 VCG16 下采样模块改进为 ChostNet, 降低了模型参数量，提高了模型准确率与检测效率;融合改进的 SegNext 注意力机制，优化了模型对较小特征的分割效果;引入改进的特征金字塔网络 (Feature pyramidnetwork, FPN) 解码架构，增强模型对语义信息的捕捉能力。其次，在检测目标割痕后，采用 Shi-Tomasi 算法，使用矩阵的最小特征值作为角点响应函数，并引入非最大值抑制筛选出制胶轨线的最佳起始点与终止点;最后，使用随机采样一致性 (Random sample consensus,RANSAC) 算法多次迭代以筛选出割痕的下边缘点，再拟合出精确的割胶轨线。试验结果表明：改进 UNet 模型的平均交并比和平均像素准确率分别为 92.86% 和 96.96%, 相比原 UNet 模型分别提高了 1.38、0.52 个百分点。此外，该模型通过结构优化将模型内存占用量压缩至 18.55MB, 较原始的 UNet 降低了 80.46%;Shi-Tomasi 角点检测算法起始点平均定位误差 10.2 像素，终止点平均定位误差 8.99 像素，可有效检测出割痕起始点和终止点;RANSAC 算法拟合割线的平均内点均方误差为 4.26 像素 2, 平均内点均方根误差为 1.86 像素。研究结果可为橡胶林复杂环境下智能化装备割痕检测与割胶轨线识别提供参考。

Abstract:

Detection of incisions on natural rubber trees and identification of rubber tapping lines are of great significance for achieving automated tapping and enhancing the intelligence of operations. In response to the problem that the recognition of incisions and the identification of tapping lines by tapping machines are easily disturbed by the natural environment under complex conditions, leading to incomplete incision recognition and inaccurate tapping line fitting, a method for detecting incisions and identifying tapping lines on rubber trees in complex environments based on an improved UNet was proposed. To address the issue of low recognition efficieney of the UNet model, the VGG16 downsampling module in the original UNet was improved to ChostNet, reducing the model's parameter quantity, improving the model's accuracy and detection efficiency; the improved SegNext attention mechanism was integrated to optimize the model's segmentation effect on smaller features; and an improved feature pyramid network (FPN) decoding architecture was introduced to enhance the model's ability to capture semantic information. Secondly, after detecting the target incision, the Shi - Tomasi algorithm was used, with the minimum eigenvalue of the matrix as the corner response function, and non-maximum suppression was introduced to select the best starting and ending points of the tapping line. Finally, the random sample consensus (RANSAC) algorithm was used for multiple iterations to select the lower edge points of the incision and then fit the precise tapping line. The experimental results showed that the average intersection over union and average pixel accuracy of the improved UNet model vere 92. 86% and 96. 96%, respectively, which were 1.38 and 0. 52 percentage points higher than those of the original UNet model. In addition, the model's memory usage was compressed to 18. 55 MB through structural optimization, a reduction of 80. 46% compared with that of the original UNet. The average positioning error of the starting point by the Shi - Tomasi cormer detection algorithm was 10. 2 pixels, and the average positioning error of the ending point was 8. 99 pixels, which can effectively detect the starting and ending points of the incision. The average inlier mean square error of the fitted line by the RANSAC algorithm was 4. 26 pixels2, and the average inlier root mean square error was 1. 86 pixels. The research results can provide a reference for the detection of incisions and identification of tapping lines in complex environments in rubber plantations using intelligent equipment.