基于FPGA加速的Mask R-CNN稻瘟病高通量自适应识别模型研究

针对基于图像的稻瘟病现场检测技术依赖先验知识且受制于算力与田间网络状况，无法实现自适应实时检测的问题，提出一种可利用现场可编程门阵列(Field programmable gate array, FPGA)加速的Mask R-CNN(Mask region-based convolutional neural network)稻瘟病高通量自适应快速识别模型。首先将骨干网络改进为MobileNetV2，利用其倒残差模块降低计算量，提高模型并行处理能力；随后增加用于稻瘟病多尺度特征融合的特征金字塔网络模块，使模型具备多尺度自适应处理能力；最后由全卷积网络(Fully convolutional network,FCN)分支输出稻瘟病病斑的实例分割，同时使用交叉熵损失函数完成稻瘟病的定位与分类。稻瘟病实测数据集对模型的验证结果表明：当输入为全高清图像时，模型平均推理时间减少至85ms，相较GPU服务器、同级别GPU边缘计算平台，速度分别提高86.2%、63.0%。在交并比为0.6时，准确率可达98.0%，病斑捕获能力平均提升21.2%。提出的Mask R-CNN自适应快速识别模型能够在田间恶劣网络状况下实现稻瘟病的快速现场检测，具有更好的抗噪能力和鲁棒性能，为水稻病害实时检测、察打一体提供了高效实时的片上系统方案。

Research on High-througput Adaptive Recognition Mask R-CNN Model for Rice Blast Disease Based on FPGA Acceleration

Image-based on-site detection technology for rice blast relies on prior knowledge which is affected by computational power and field network conditions, rendering adaptive real-time detection impossible. To tackle these challenges, a Mask R-CNN (Mask region-based convolutional neural network) model for rapid, high-throughput, and adaptive identification of rice blast was proposed. This model can be accelerated by using field programmable gate array (FPGA). Firstly, the backbone network was replaced with MobileNetV2, leveraging its inverted residual module to decrease computations and enhance the model’s parallel processing capabilities. Following that, a feature pyramid network module was incorporated to facilitate multi-scale feature fusion for rice blast, enabling the model to possess multi-scale adaptive processing abilities. Finally, the fully convolutional network(FCN) branch outputed the instance segmentation of rice blast lesions, utilizing the Softmax function to accurately localize and classify rice blast diseases. The validation results of the model using test datasets for rice blast disease demonstrated significant capabilities: when the input was a full HD image, the average inference time of the model was reduced to 85ms, which was 86.2% and 63.0% faster than the GPU server and the same level GPU edge computing platform, respectively. When the intersection over union ratio was 0.6, the accuracy can reach 98.0%, and the disease spot capture ability was improved by 21.2% on average. The Mask R-CNN adaptive fast identification model proposedcan realize the rapid field detection of rice blast disease under severe network conditions, and had better anti-noise ability and robust performance, which provided an efficient real-time system-on-chip scheme for real-time detection, inspection and mitigation of rice disease.