基于改进YOLOv3-Tiny的番茄苗分级检测

Tomato seedling classification detection using improved YOLOv3-Tiny

为了提高番茄苗分选移栽分级检测精度,该研究提出了YOLOv3-Tiny目标检测改进模型。首先建立了番茄穴盘苗数据集,使用K-means++算法重新生成数据集锚定框,提高网络收敛速度和特征提取能力;其次为目标检测模型添加SPP空间金字塔池化,将穴孔局部和整体特征融合,提高了对弱苗的召回率;同时加入路径聚合网络(PANet),提升细粒度检测能力;引入了SAM空间注意力机制,提高对番茄苗的关注,减少背景干扰;增加了ASFF(Adaptively Spatial Feature Fusion)自适应特征融合网络,能够使目标检测模型对多个级别的特征进行空间滤波;采用CIoU损失函数策略,提高模型收敛效果。改进的YOLOv3-Tiny目标检测模型经过数据集训练,在测试集上能够达到平均精度均值为97.64%,相比YOLOv3-Tiny模型提高了3.47个百分点。消融试验验证了网络结构改进和训练策略是有效的,并将改进的YOLOv3-Tiny目标检测算法与5种目标检测算法进行对比,发现改进的YOLOv3-Tiny目标检测模型在重合度阈值为50%的条件下平均精度均值为97.64%。单张图像处理时间为5.03ms,较其他目标检测算法具有明显的优势,验证了该模型能够满足番茄苗分级检测精度要求,可以为幼苗分选检测方法提供参考。

Here, an improved YOLOv3-Tiny target detection model was proposed to enhance the detection accuracy of the seedling classification and detection in the process of tomato seedling transplanting. First of all, 2 160 images of the individual tomato seedling were collected to pre-process the image of the tomato hole. A LabelImg software was then selected to mark the image. After that, the data enhancement was performed on the image, such as the rotate and flip operations. As such, 25 080images were generated, where 22 800 images were taken as the training set, and 2280 images were the test set. A target detection model of tomato hole seedling was improved for the better convergence speed of the network and the feature extraction, where the K-means ++ was used to regenerate the anchors of the tomato plug seedling dataset. Secondly, a Spatial Pyramid Pooling(SPP) was added into the target detection model, further to integrate the local and global features of the plug holes for the less recall rate of weak seedlings. A path aggregation network(PANet) was also added to improve the fine-grained detection. A spatial attention mechanism(SAM) was then introduced to reduce the background noise in the target detection model. An adaptive feature fusion network was selected to directly learn the features from the other levels, where spatial filtering was performed for better features fusion. A CIoU loss function strategy was adopted to improve the convergence of the model. Eventually, the model training was conducted in a computer-deep learning environment after the dataset production and network construction. The results show that the Mean Average Precision value reached 97.64%, which was higher than 94.17% of the original. The F1 value of the improved YOLOv3-Tiny reached 0.94, which was higher than 0.92 of the original. A comparative experiment was also performed on the different types of tomato plug seedlings, further to verify the effectiveness and feasibility of the improved model. It was found that the improved YOLOv3-Tiny target detection model was fully met the requirements of tomato plug seedling grading detection, where Average Precision values were 98.22%,94.69%, and 99.99% for the strong, weak, and no seedlings, respectively. Additionally, the improved network structure and the training strategy were used to verify the model in the process of the ablation experiment. We found that every improvement method of the model in the research has positive significance, and the introduction of PANet has the most obvious improvement in the Mean Average Precision value of the model, which is increased by 2.17 percentage points. Using the improved YOLOv3-Tiny target detection algorithm to compare with target detection algorithms such as YOLOv3-Tiny,Faster-RCNN and CenterNet, it is found that under the condition that the overlap threshold is 50%, the Mean Average Precision of the improved YOLOv3-Tiny The value is still 0.47 percentage points higher than other CenterNet algorithms with the highest Mean Average Precision value;the improved YOLOv3-Tiny detection time is 5.03 ms per image, which is 8.39 ms less than the YOLOv3 large target detection algorithm with the shortest detection time. The finding can also provide a strong reference to detect the seedling sorting during tomato production.