注意力特征融合的番茄叶部早期病斑诊断算法

Early lesion diagnosis algorithm of tomato leaf based on attention feature fusion

番茄产量受到病害、天气等因素的影响,其中番茄生长过程中叶片的病害问题是影响番茄产量的最关键因素。然而,在叶片病害检测领域,现有模型普遍存在泛化能力不足以及小病斑漏检率高等问题。提出一种改进的番茄病害早期检测算法,通过对YOLOv5s网络进行多方面的优化来改善这些问题,同时保持模型轻量化。首先,采用Mosaic 9数据增强技术,强化了模型对小病斑的检测能力,增加了图像背景的复杂度,提高了模型的泛化能力;其次,使用GSConv和Slim-Neck网络,在保持模型准确性的前提下轻量化模型,降低计算负担;同时,使用SimAM注意力机制更准确地捕捉叶片上的小病斑特征,从而降低漏检率;此外,为了进一步增强多尺度目标的检测能力,引入自适应空间特征融合,有效地整合不同尺度的特征,提升了多尺度目标,特别是小目标的检测准确性。实验结果表明:该模型在叶片病害早期检测方面表现出色,对叶霉、早疫、晚疫以及健康叶片四种番茄病害的早期平均识别准确率、召回率、F1分数及mAP分别达到了0.951%、0.918%、0.934%、0.948%。可见该方法对于小病斑具有较好的检测性能,改善了模型泛化能力不足及小病斑检测过程中的漏检问题,进一步提高了检测效果。

Tomato yield is affected by diseases,weather and other factors,among which leaf disease is the most critical factor affecting tomato yield.However,in the field of leaf disease detection,the existing models generally have the problem of insufficient generalization ability and high detection rate of small lesions.In this paper,an improved tomato disease early detection algorithm is proposed to improve these problems by optimizing the YOLOv5s network in various aspects,while keeping the model lightweight.Firstly,Mosaic9 data enhancement technology is used to strengthen the detection ability of the model for minor lesions,increase the complexity of the image back-ground,and improve the generalization ability of the model.Secondly,GSConv and Slim-Neck networks are used to lightweight the model and reduce the computational burden while maintaining the accuracy of the model.At the same time,the SimAM attention mechanism was used to capture the features of small lesions on the leaves more accurately,thus reducing the missed detection rate.In addition,in order to further enhance the detection ability of multi-scale targets,adaptive spatial feature fusion is introduced to effectively integrate features of different scales,and improve the detection accuracy of multi-scale targets,especially small targets.The experimental results showed that the model had excellent perfor-mance in early detection of leaf diseases,and the average recognition accuracy,recall rate,F1 score and mAP of leaf mold,early disease,late disease and healthy leaf disease reached 0.951%,0.918%,0.934%and 0.948%,respectively.It can be seen that this method has a good detection performance for minor lesions,and improves the problem of insufficient generalization ability of the model and missing detection in the detection process of minor lesions,and further improves the detection effect.