基于GGWL-YOLO模型的轻量化熟草莓识别方法

Recognition Method of Lightweight Ripe Strawberry Based on GGWL-YOLO Model

摘要:

由于草莓果实受枝叶遮挡以及光线变化的复杂背景影响，草莓果实识别精度较低，影响自动化采摘效率。为实现复杂环境下的草莓果实目标快速识别，本文提出基于YOLO 11改进的GGWL-YOLO成熟草莓识别模型。首先，引入多层次特征提取的HGNetV2网络，并加入多尺度池化技术的SPPF模块和结合多头自注意力与前馈神经网络的C2PSA模块，构建GhostHGNetV2网络作为特征提取主干网络，适用于草莓果实检测，更进一步提取草莓果实的未遮挡部分细小特征。其次，在颈部网络引入GFPN网络，用DySample改进上采样模块，构建GDFPN网络实现多层次跨尺度特征融合，与GhostHGNetV2网络细小特征提取能力相结合，改善草莓果实识别的遮挡问题。然后，选用WIoUv3改进边界框回归损失函数，改善因光照影响草莓果实识别问题，减少草莓图像低质量样本的影响。最后，使用Lamp方法对改进的整体模型进行通道剪枝和模型微调，减少冗余通道数保留关键的通道信息。试验结果表明，改进的GGWL-YOLO模型的精确率、mAP@0.5、参数量、检测速度分别为92.5%、94.3%、3.28×10^6、91.5 f/s，与YOLO v8s、YOLO 11s、RT-DETR-r18模型相比精确率分别提升1.4、1.2、2.4个百分点，参数量减少70.4%、65.1%、83.4%，并且在平均精度均值和浮点运算量也存在优势。改进后的模型可以快速、准确地识别复杂环境下的草莓果实，进一步推动草莓自动化采摘。

Abstract:

Due to the influence of the complex background of strawberry fruit and branches and leaves as well as light changes, the recognition accuracy of strawberry fruit is low. It affects the efficiency of automated picking. To achieve rapid recognition of strawberry fruit targets in complex environments, an improved GGWL-YOLO mature strawberry recognition model based on YOLO 11 was proposed. Firstly, HGNetV2 network with multi-level feature extraction was introduced, SPPF module with multi-scale pooling technology and C2PSA module with multi-head self-attention and feedforward neural network were added. GhostHGNetV2 network was constructed as the feature extraction backbone network, which was suitable for strawberry fruit detection. Further the fine features of the uncovered part of strawberry fruit was extracted. Secondly, the GFPN network was introduced into the neck network, the upsampling module was improved with DySample, the GDFPN network was constructed to achieve multi-level and cross-scale feature fusion, and it was combined with the fine feature extraction ability of the GhostHGNetV2 network to improve the occlusion problem in strawberry fruit recognition.Then, WIoUv3 was used to improve the bounding box regression loss function, improve the problem of strawberry fruit recognition due to the influence of light, and reduce the influence of low-quality strawberry picture samples. Finally, Lamp method was used for channel pruning and model fine-tuning of the improved overall model to reduce the number of redundant channels and retain key channel information. The test results showed that the precision, mAP@0.5, parameter number and detection speed of the improved GGWL-YOLO model were 92.5%, 94.3%, 3.28×10^6 and 91.5 f/s, respectively. Compared with YOLO v8s, YOLO 11s and RT-DETR-r18 models, the precision was improved by 1.4, 1.2 and 2.4 percentage points, respectively, while the number of parameters was reduced by 70.4%, 65.1% and 83.4%. Moreover, it also had advantages in mean average accuracy and floating point computation. The improved model can quickly and accurately identify strawberry fruits in complex environments, and further promote automatic strawberry picking.