基于YOLO v8n-seg与轻量化CNN的草莓成熟度级联检测方法

Strawberry Ripeness Cascade Detection Method Based on YOLO v8n-seg and Lightweight CNN

摘要:

针对自然环境下草莓成熟度识别受枝叶遮挡及光照不均影响导致分类精度低的问题,本研究提出一种融合改进的YOLO v8 实例分割与轻量化CNN分类网络的级联检测方法。该方法首先使用GhostNet轻量化结构和注意力机制改进的YOLO v8n-seg模型实现果实定位与分割,在模型参数量较低(2. 6 × 106)的前提下,实现在复杂背景中对目标的准确分割。其次,利用分割掩码从原始图像中裁剪出果实目标区域,并输入至基于EfficientNet-Lite构建的CNN分类网络,该网络借助MBConv(Mobile inverted bottleneck convolution)模块和Softmax 分类器,专注于捕捉果实表面细微的颜色与纹理特征变化,有效提升了模型对“转熟”等过渡阶段果实的辨识能力。试验结果表明, 本研究提出的解耦式空间定位与属性识别协同策略显著增强了系统的鲁棒性,成熟度识别平均准确率达到 88. 1% ,整体系统参数量为7. 3 × 106(EfficientNet-Lite参数量为4. 7 × 106),推理速度为71. 9 f/ s,在保证实时处理的同时实现了对未熟、转熟及成熟草莓的准确分级,为草莓智能采摘机器人的视觉系统提供了技术支撑。

Abstract:

Aiming to address the problem of low classification accuracy in strawberry ripeness recognition caused by branch and leaf occlusion and uneven lighting under natural conditions, a cascade detection method that integrated an improved YOLO v8 instance segmentation with a lightweight CNN classification network was proposed. This method firstly used the GhostNet lightweight structure and an attention mechanism to improve the YOLO v8n-seg model for fruit localization and segmentation, achieving accurate target segmentation in complex backgrounds with a relatively low model parameter count (2. 6 × 106). Nextly, fruit target areas were cropped from the original images by using the segmentation masks and input into a CNN classification network built on EfficientNet-Lite. This network, with mobile inverted bottleneck convolution (MBConv) modules and a Softmax classifier, focusing on capturing subtle color and texture feature changes on the fruit surface, effectively improved the model's ability to recognize fruits in transitional ripening stages. Experimental results showed that the proposed decoupled spatial localization and attribute recognition coordination strategy significantly enhanced system robustness. The average ripeness recognition accuracy reached 88. 1% , the overall system parameter count was 7. 3 × 106 (EfficientNet-Lite parameters accounted for 4. 7 × 106 ), and the inference speed was 71. 9 f/ s, achieving accurate grading of unripe, turning, and ripe strawberries while ensuring real-time processing. The research result can provide technical support for the vision system of intelligent strawberry-picking robots.