视觉注意与协同学耦合机制下的农田杂草图像分割及检测

Although image segmentation may be considered as the currently major step for detecting the objects using machine vision, its instability resulting from variability of outdoor operating environment is a main obstacle to influence directly the accuracy of object detection. Theories and techniques of bionics, biovision, synergetics and machine vision were employed to study the main drycrop, including rape, cotton and corn, in middle and lower Yangtze river. Aiming at quick access to the precision location of crop,visual attention model in this study was firstly proposed to locate region-of-interest, which solved the obstacle of traditional image segmentation methods, on the premise that accuracy and real-time requirements of field work. Then, inter-row and intra-row weeds in fields were detected using coupling mechanism of visual attention and synergetics and features of crop cultivation. Throught analyzing the results from image segmentation experiments based-on traditional algorithm and visual attention model, research how to influence image segmentation by variability of illumination and background in outdoor operating environment was done using synergetic perception theory and image performance evaluation. From above analysis, the mechanisms of enhancing the performance of image segmentation using visual attention model was explicated in order to discover the universal algorithm for detecting weeds. The due results are obtained to expand applied fields of visual attention mechanism and machine vision and to generate novel idea and method of site-specific weed management.

图像分割是现阶段运用机器视觉进行目标检测的主要步骤，但其技术瓶颈在于传统图像分割方法易受到外界环境干扰存在不稳定性，直接影响目标检测精度。本项目以长江中下游3种旱地农作物（油菜、棉花和玉米）为研究对象，综合运用仿生学、生物视觉、协同学及机器视觉理论和技术，在满足田间作业实时性和准确性前提下，以快速准确定位作物区域为目标，解决图像分割不稳定性问题，突破传统图像分割瓶颈，进而在视觉注意和协同学耦合作用下结合作物栽培特点，实现田间环境下行间和株间杂草快速准确检测。运用协同感知理论和图像性能评价技术，结合传统图像分割和视觉注意模型下图像分割对比试验结果，分析光线干扰和背景干扰对图像分割的影响，总结归纳视觉注意模型下图像分割性能改善机理，探索通用性杂草检测算法。预期研究成果将拓展视觉注意机制和机器视觉的应用领域，为田间杂草定位作业提供新的思路和方法。

现阶段，田间杂草作业中植株定位是精准作业环节之一。根据视觉注意理论能够快速筛选目标的特性，本项目提出了一种新的杂草检测方法，即利用视觉注意和协同学理论结合大田作物栽培特性建立相应数学模型，提取感兴趣区域显著图，直接分割出植株图像，而不受外界因素如光线和背景变化干扰，改善传统图像分割易受外界环境干扰存在不稳定性直接影响目标检测精度问题，进而反演出杂草区域。本项目从视觉注意模型建立、图像分割性能改善机理、探索通用性杂草检测算法等思路进行了研究。结果显示：（1）视觉注意理论能够准确聚焦感兴趣区域，有望突破传统图像分割瓶颈。结合作物颜色特性以及田间分布特性，分割效果以及准确度强于单一颜色指标，正确识别率可达95.09%，误分割率达1.51%，漏分割率为4.91%。（2）由于视觉注意模型和协同感知理论耦合作用机制，很大程度上避免光线分布不均以及背景复杂变化（如土壤不同含水率、残茬覆盖等）带来的图像分割不完整问题，改善传统图像分割机理，结合图像性能评价方法，进一步从技术层面明确提出的分割模型对改善机理的定量和定性表征。（3）通用性杂草检测算法研究表明，视觉注意模型可有效分割不同品种植株进而识别行间和株间杂草区域。