基于深度学习和视觉显著性算法的自然环境中烟叶成熟度判别

Maturity Discrimination of Tobacco Leaves in Natural Environments Based on Deep Learning and Visual Saliency Algorithm

摘要:

自然环境中烟叶成熟度不同步的特性对选择性采摘提出了挑战，现有的烟叶成熟度判别算法在复杂光照条件下鲁棒性不足，且未能有效平衡轻量化程度和分类精度。本研究提出一种融合视觉显著性算法与自适应光照校正的深度学习方法，以实现田间烟叶成熟度的精准判别。使用经过轻量化的YOLO v8n检测图像中的烟叶区域；其次，对检测出的烟叶区域使用自适应Gamma校正，减少光照变化对成熟度判别的影响；利用改进的视觉显著性算法生成烟叶的显著性图；利用改进MobileNetV2网络对烟叶成熟度等级进行判别。试验结果表明，轻量化后的YOLO v8n检测模型内存占用量为原模型的29%，平均精度均值达94.8%。在输入图像为RGB三通道时所提方法的判别准确率为91.93%，加入显著性增强算法后，准确率达到95.07%，比目前最好的分类模型(MobileNetV3Large)高2.03个百分点。加入自适应Gamma校正算法后，高光照情况下烟叶图像信息熵提升7.5%。结果证明了所提出的自适应Gamma校正与显著性增强结合的烟叶成熟度判别方法的有效性，可为烟叶采摘机器人提供技术支持。

Abstract:

The asynchronous maturity of tobacco leaves in natural environments presents a significant challenge for selective harvesting. Existing algorithms for tobacco leaf maturity classification often lack robustness under complex lighting conditions and struggle to balance model light weightness with classification accuracy. To address these issues, a deep learning-based approach that integrated a visual saliency algorithm with adaptive illumination correction to achieve accurate maturity classification of tobacco leaves in field environments was proposed. Firstly, the lightweight YOLO v8n model was used to detect tobacco leaf regions in the input image. Then, adaptive Gamma correction was applied to these regions to mitigate the impact of varying illumination on maturity classification. Next, an improved visual saliency algorithm was employed to generate saliency maps of the tobacco leaves. Finally, an enhanced MobileNetV2 network was used to classify the maturity levels of the detected leaves. Experimental results demonstrated that the lightweight YOLO v8n model reduced memory usage to 29% of the original model while achieving a mean average precision (mAP@0.5) of 94.8%. When using RGB images as input, the proposed method attained a classification accuracy of 91.93%. After integrating the saliency enhancement algorithm, the accuracy was increased to 95.07%, outperforming the current state-of-the-art model (MobileNetV3Large) by 2.03 percentage points. Additionally, the use of adaptive Gamma correction increased the information entropy of tobacco leaf images by 7.5% under high illumination conditions. These results validated the effectiveness of the proposed method, which combined adaptive Gamma correction with visual saliency enhancement. The approach offered a promising solution for improving tobacco leaf maturity classification and provided technical support for the development of tobacco harvesting robots.