基于YOLO v7-RA的火龙果品质与成熟度双指标检测方法

Dual-index Detection Method of Pitaya Quality and Maturity Based on YOLO v7-RA

摘要:

已有火龙果检测方法仅针对单一性能指标，难以满足农业真实场景的需要，为此提出了一种精准高效的火龙果品质与成熟度双指标检测方法。首先，利用自适应鉴别器增强的样式生成对抗网络扩充火龙果图像，建立复杂环境火龙果数据集。采用伽马变换进行图像增强，凸显火龙果特征，降低光照环境的影响。其次，提出了YOLO v7-RA模型。通过设计ELAN_R3替代ELAN（Efficient layer aggregation network）模块，减少主干网络对重复特征的提取，增强模型对细粒度特征关注度，提高双指标检测准确率。融入混合注意力机制（Mixture of self-attention and convolution，ACmix），增强模型对特征的提取和整合能力，降低杂乱背景信息干扰。最后，通过实验验证了YOLO v7-RA模型的检测性能。实验结果表明，该方法精准率为97.4%，召回率为97.7%，mAP0.5为96.2%，FSP为74f/s，实现了检测精度与检测速度的均衡。即使在遮挡情况下，YOLO v7-RA模型检测精准率仍达到91.4%，具有较好泛化能力，能够为火龙果智能化采摘的发展提供技术支持。

Abstract:

Research on pitaya detection methods is the basis and prerequisite for realizing intelligent picking. Existing pitaya detection methods only target a single performance indicator, which is difficult to meet the needs of real agricultural scenarios. Therefore, an accurate and efficient dual-index detection method for pitaya quality and maturity was proposed. Firstly, the adaptive discriminator enhanced style generation adversarial network algorithm was used to expand the pitaya image and establish a pitaya dataset. The image was enhanced by gamma transform to highlight the characteristics of pitaya and reduce the impact of lighting environment. Secondly, the YOLO v7-RA model was proposed, by designing ELAN_R3 to replace the efficient layer aggregation network (ELAN) module to reduce the extraction of repetitive features by the backbone network. This enhanced the model’s attention to fine-grained features and improved the accuracy of dual-index detection. The mixture of selfattention and convolution (ACmix)was applied to enhance the model’s ability to extract and integrate feature information, and reduce the interference of cluttered background information. Finally, the detection level of the YOLO v7-RA model was verified through experiments. Experimental results showed that the precision rate of the method was 97.4%, the recall rate was 97.7%, the mAP0.5 was 96.2%, and FSP was 74f/s. A balance between detection accuracy and detection speed was achieved. Even under occlusion, the YOLO v7-RA model detection accuracy still reached 91.4%. The model had good generalization ability to provide strong technical support for the development of intelligent pitaya picking.