【摘要】 利用具有强氧化性、无毒无污染的臭氧水，借助工厂化、无人化处理设备和工艺，根据果蔬种类和农药残留状况，喷淋不同浓度的臭氧水，是实现果蔬农药残留自动化、智能化、绿色、高效消解处理的有效手段。而快速、精准识别果蔬种类是指导工厂化、无人化处理设备针对性喷淋所需浓度臭氧水的前提。本文提出一种基于改进YOLO v7-tiny的农残消解过程中多品种果蔬精准识别方法。首先，在Head部分引入渐近特征金字塔网络AFPN（Asymptotic feature pyramid network）来替换原有YOLO v7-tiny的金字塔网络，在实现网络结构轻量化的同时进一步提高模型准确率；其次，在Backbone部分加入多尺度注意力模块EMA（Efficient multi-scale attention）来提高模型对果蔬图像中有效特征信息的提取能力；最后，将原有YOLO v7-tiny网络的损失函数替换为Wise-IoU（Wise intersection over union）以提高模型的泛化能力。试验结果表明，改进后YOLO v7-tiny模型（YOLO v7-AEW）的参数量、浮点数计算量和模型存储占用量分别达到4.5×10⁶、1.28×10¹⁰和8.9 MB,较原模型分别减少25.0%、3.0%和27.6%;准确率与平均精度均值达到97.9%和96.8%,较原模型分别提高2.3、1.3个百分点；改进后模型与Faster R-CNN、SSD、YOLO v5s、YOLO v8算法对比，平均识别精度均值分别提高37.1、38.0、32.5、7.1个百分点。搭建基于改进YOLO v7-tiny的农残消解装置并进行检测试验，改进后模型的检出率和漏检率分别为97%和3%,检测效果优于其他网络模型，证明改进后模型具有较高的实际应用价值。本研究可为基于臭氧水的工厂化、无人化果蔬农药残留消解一体机研制提供深度学习识别方法参考。更多还原

【Abstract】 Utilizing ozone water with strong oxidation property, which is healthy and pollution-free, spraying ozone water in accordance with different types of fruits and vegetables and diverse pesticide residues is an effective approach to achieve automatic, intelligent, green, and efficient digestion treatment of fruit and vegetable residues. The rapid and precise identification of fruit and vegetable types is the prerequisite for guiding the concentration of ozone water required for targeted spraying of factory and unmanned treatment equipment. A method based on the improved YOLO v7-tiny was proposed to accurately identify multiple varieties of fruits and vegetables during the process of agricultural residues digestion. Firstly, the asymptotic feature pyramid network（AFPN） was introduced in the Head part to substitute the original YOLO v7-tiny pyramid network. Secondly, the efficient multi-scale attention module（EMA） was added to the backbone to enhance the model’s capability of extracting effective feature information from fruit and vegetable images. Finally, the loss function of the original YOLO v7-tiny network was replaced by Wise-IoU. The experimental results indicated that the parameter number, computational capacity, and model size of the improved YOLO v7-tiny model reached 4.5×10⁶, 1.28×10¹⁰, and 8.9 MB respectively, which were decreased by 25.0%, 3.0%, and 27.6% compared with that of the original model. The average accuracy and average precision were 97.9% and 96.8%, which were 2.3 percentage points and 1.3 percentage points higher than those of the original model respectively. Compared with Faster-RCNN, SSD, YOLO v5s, and YOLO v8 algorithms, the average recognition accuracy of the improved model was increased by 37.1 percentage points, 38.0 percentage points, 32.5 percentage points, and 7.1 percentage points respectively. An agricultural residue digestion device based on the improved YOLO v7-tiny was constructed and the detection test was conducted. The detection rate and missing rate of the improved model were 97% and 3% respectively, and the detection effect was superior to other network models, which proved that the improved model had high practical application value. The research result would provide a reference for the deep learning identification method for the development of an integrated machine for industrial and unmanned fruit and vegetable residues digestion based on ozone water.更多还原