基于实例分割的玉米虫害检测研究

Research on Maize Pest Detection Based on Instance Segmentation

为实现玉米虫害精准、快速识别,提出基于实例分割的卷积神经网络结合迁移学习的虫害检测方法。以草地贪夜蛾的卵、幼虫、成虫为检测对象,使用数据增强对图像数据进行扩充,将Yolact++模型在COCO数据集上的预训练权重迁移至草地贪夜蛾的检测。使用focal loss中解决难易样本不均衡的思想,优化模型中的损失函数。以Detnet模型改进Yolact++模型中Resnet主干模型部分,提高对小目标检测的效果。训练过程中使用卷积层先冻结再解冻、局部与全局相结合的训练方法,获得最优权重模型,并对模型进行测试。结果表明:该方法与YoloV3+迁移学习、SSD+迁移学习、Yolact+迁移学习、Yolact++等检测算法相比,对复杂背景图像检测有更好的准确率与漏检率,测试的准确率达到96.32%,漏检率为5.51%,误检率为5.33%。

In order to achieve accurate and rapid identification of maize pests,this paper proposes a pest detection method using convolutional neural network combined with transfer learning based on instance segmentation. Taking eggs,larvae and adults of Spodoptera frugiperda as detection objects,the image data was expanded by data enhancement,and the pre-training weights of Yolact++ model on COCO data set were migrated to the detection of Spodoptera frugiperda. Using the idea of focal loss to solve the imbalance of difficult and easy samples,the loss function in the model was optimized. The Detnet model was used to improve the Resnet trunk model in the Yolact++ model to improve the effect of small target detection. In the training process,the convolution layer was first frozen and then thawed,and the local and global training methods were combined to obtain the optimal weight model,and the model was tested. The test results showed that compared with the detection algorithms of YoloV3+migration learning,SSD+migration learning,Yolact+migration learning and Yolact++,this method had better accuracy and missed detection rate for complex background image detection. The accuracy of the test reached 96. 32%,the missed detection rate was 5. 51%,and the false detection rate was 5. 33%.