基于改进YOLO v8的轻量化稻瘟病孢子检测方法

Lightweight Rice Blast Spores Detection Method Based on Improved YOLO v8

摘要:

稻瘟病由稻瘟病孢子通过空气进行传播，严重影响水稻产量，因此，稻瘟病孢子的检测对于稻瘟病早期诊断与防治具有重要作用。针对现有方法存在检测速度慢的问题，本研究基于YOLO v8模型提出了一种稻瘟病孢子检测方法RBS-YOLO。首先，该算法在主干网络中引入PP-LCNet轻量化网络结构，减少模型每秒浮点运算次数并降低模型内存占用量，其次在颈部网络中引入高效多尺度注意力模块(Efficient multiscale attention module, EMA)，并将原损失函数改进为WIOU损失函数，提高了模型识别稻瘟病孢子的精确率与平均精度均值。改进后的RBS-YOLO模型精确率与平均精度均值分别为97.3%和98.7%，满足稻瘟病孢子的检测需求，模型内存占用量与每秒浮点运算次数分别为3.46MB、5.2×109，同YOLO v8n相比分别降低41.8%与35.8%。RBS-YOLO模型与当前主流的YOLO v5s、YOLO v7、YOLO v8n模型对比,每秒浮点运算次数分别降低67.3%、95.1%、35.8%。研究结果表明RBS-YOLO模型能够满足稻瘟病孢子实时检测的需求，且有利于部署到移动端。

Abstract:

Rice blast is one of the most serious diseases of rice. It is caused by blast fungus and occurs in different growth stages of rice. The spores of blast can be transmitted through air, which seriously affects food production security. Therefore, the identification of blast spores plays an important role in the early diagnosis and control of rice blast. Based on the YOLO v8 model, an RBS-YOLO method for the detection of rice blast spores was proposed. Firstly, the algorithm introduced the PP-LCNet lightweight network in the backbone network, which used DepthSepConv as the basic block and reduced the computational effort of the model and the size of the model weight file, but hardly increased the inference time. Secondly, the efficient multi-scale attention module was introduced into the neck network, which reshaped some channels into batch dimensions and grouped the channel dimensions into multiple sub-features, so that the spatial semantic features were evenly distributed in each feature group. The information of each channel can be effectively preserved and the computational overhead can be reduced. Finally, the loss function of YOLO v8n was changed to WIOU loss function, which can reduce the impact of low-quality samples on the model during training. WIOU used dynamic non-monootone focusing mechanism to evaluate the quality of the anchor frame, and used gradient gain, which ensured the high-quality effect of the anchor frame and reduced the influence of harmful gradients. The accuracy and mean accuracy of model identification of rice blast spores were improved. The accuracy and average accuracy of the improved RBS-YOLO model were 97.3% and 98.7%, respectively, meeting the demand for the detection of rice blast spores. The weight file size and computation amount were 3.46MB and 5.2×109, respectively, which were 41.8% and 35.8% lower than that of YOLO v8n. In order to verify the detection performance of RBS-YOLO, under the same training environment and parameter configuration, the improved model was compared with the YOLO v5s, YOLO v7 and the original YOLO v8n model, and the computational load was reduced by 67.3%, 95.1% and 35.8%, respectively. Model weight file sizes were reduced by 10.14MB, 67.84MB, and 2.49MB, respectively. The results showed that RBS-YOLO can meet the demand of real-time detection of rice blast spores,which was conducive to deployment to mobile terminals.