基于改进YOLOv5的菇房平菇目标检测与分类研究

随着食用菌行业由自动化向智能化、信息化发展的趋势越来越明显，为了实现现代化菇房中平菇的准确检测，解决工厂化平菇栽培中收获阶段平菇之间相互遮挡等问题，帮助平菇采收机器人进行准确的自动化采收，该文提出了一种基于YOLOv5（you only look once version 5）模型的OMM-YOLO（ostreatus measure modle-YOLO）平菇目标检测与分类模型。通过在YOLOv5模型的Backbone层添加注意力模块，对输入的平菇图像特征进行动态加权，以获得更详细的特征信息，并在Neck层采用加权双向特征金字塔网络，通过与不同的特征层融合，提高算法的平菇目标检测的精度。此外，为了改善算法的准确性和边界框纵横比的收敛速度，该文采用了EIoU（enhanced intersection over union）损失函数替代了原有的损失函数。试验结果表明，与原始模型相比，改进模型OMM-YOLO对成熟平菇、未成熟平菇和未生长平菇的平均精度均值分别提高了0.4个百分点、4.5个百分点和1.1个百分点。与当前主流模型Resnet50、VGG16、YOLOv3、YOLOv4、YOLOv5m和YOLOv7相比，该模型的精确率、召回率和检测精度均处于优势，适用于收集现代化菇房中的平菇信息，有效避免了平菇之间因相互遮挡而产生的误检测现象。菇房平菇目标检测可以自动化地检测平菇的数量、生长状态等信息，帮助菇房工作人员掌握菇房内的菇况，及时调整温湿度等环境条件，提高生产效率，并且对可以对平菇进行质量控制，确保平菇产品的统一性和品质稳定性。同时可以减少对人工的依赖，降低人力成本，实现可持续发展，对智能化现代菇房建设具有积极作用。

Object detection and classification of pleurotus ostreatus using improved YOLOv5

Accurate detection of pleurotus ostreatus is highly required in modern mushroom houses, particularly with the ever-increasing automation of intelligence and informatization. However, manual harvesting cannot fully meet the large-scale factory cultivation in the edible mushroom industry. Fortunately, a harvesting robot can be expected to increase productivity with the reduced labor intensity at the industrial cultivation and harvesting stage. In this article, an improved OMM-YOLO(ostreatus measure modle-you only look once) target detection model was proposed to conduct comparative experiments on the mushroom detection dataset. 1353 images of mushrooms were collected from the mushroom houses and then classified into the mature, immature, and ungrown at the growth stages. An OMM-YOLO target detection and classification model was proposed using the improved YOLOv5 model. The ECA (effective channel attention) attention module was also introduced into the Backbone layer of the original YOLOv5 model. The features of input mushroom images were dynamically weighted to obtain more feature information. At the same time, a weighted bidirectional feature pyramid network (BiFPN) was used in the Neck layer. Multi-scale features were repeatedly fused from the top to the bottom and then bottom-up, thereby improving the accuracy of mushroom target detection. In addition, the EIoU (enhanced intersection over union) Loss function was utilized to improve the accuracy of the model and the rate of convergence of the bounding box aspect ratio, instead of the original Loss function CIoU (Complete IoU). The better performance was achieved in terms of the average accuracy mAP, and the rate of convergence of the bounding box aspect ratio. The experimental results show that the improved OMM-YOLO model shared an average accuracy mAP of 91.4%, 87.1%, and 95.1% for the mature, immature, and ungrown mushrooms, respectively, which was improved by 0.4, 4.5, and 1.1 percentage points, respectively, compared with the original. Better performance was achieved in the improved OMM-YOLO model, in terms of the accuracy, recall, detection accuracy, and detection speed, compared with the current mainstream models, such as Resnet50, VGG16, YOLOv3, YOLOv4, YOLOv5m, and YOLOv7. Therefore, this improved model was very suitable for the detection of mushrooms in modern mushroom houses. Mushroom features were collected to effectively avoid the occurrence of false detection caused by the mutual occlusion of mushrooms. An attention mechanism was introduced to weigh the BiFPN using a more appropriate Loss function. The accuracy and rate of convergence of the improved model were significantly enhanced for the mushroom target detection. A more effective technical means can be expected in the detection of mushroom targets in modern mushroom houses at present. As such, mushroom target detection can automatically detect the quantity and growth condition of the mushroom in the mushroom room. Environmental parameters (such as temperature and humidity) can be adjusted in a timely manner. The production efficiency can be promoted to control the quality of the mushroom for better uniformity and quality stability of the products. At the same time, it can be used to reduce the dependence on labor and costs. Sustainable development can be realized through the positive impact of the construction of intelligent mushroom houses.