Abstract: Accurate prediction of baking temperature and humidity for real-time tobacco baking images in baking room is the key to the design of intelligent baking room, the use of spatial domain features of tobacco in baking images for temperature and humidity prediction has achieved certain results, but there are still problems such as insufficient extraction of texture details of tobacco and ineffective fusion of features in the frequency domain, etc., so a tobacco baking temperature and humidity prediction model based on global-local feature fusion network in image space-frequency domain (SFGLN) was proposed . Firstly, based on the Transformer architecture, global spatial-frequency domain information and local block texture features were fused. Global space-frequency domain information referred to the extraction of global null-frequency domain tobacco information from the whole image by deep network and the extraction of global frequency domain information by Fourier transform, while local block texture features referred to the extraction of texture details of local tobacco from local blocks of the image. Finally, the fine-grained information of the local blocks was fused with the global contextual information through the attention mechanism for tobacco bakingg temperature and humidity prediction. The experimental results show that SFGLN model improves temperature prediction accuracy by 4.76% within 2 degrees and humidity prediction accuracy by 0.49% within 1 degree based on tobacco baking images only compared to the model with sub-optimal performance, which shows that the proposed model has better performance in tobacco baking temperature and humidity prediction.

Key words: temperature and humidity prediction, tobacco baking, image recognition, convolutional neural networks, feature fusion

摘要： 针对烤房中烟叶实时烘烤图像准确预测烤房烘烤温度和湿度是智能烤房设计的关键，利用烘烤图像烟叶空间域特征进行温湿度预测取得一定的效果, 但是仍存在烟叶纹理细节提取不够充分，频率域特征未能有效融合等问题, 提出一种基于图像空频域全局与局部特征融合网络（SFGLN）的烟叶烘烤温湿度预测模型。首先基于Transformer架构，融合全局空频域信息和局部块纹理特征。全局空频域信息指通过深度网络从整个图像中提取全局空域烟叶信息和采用傅里叶变换提取全局频域信息，而局部块纹理特征指从图像局部块中提取局部烟叶的纹理细节。最后通过注意力机制，将局部块的细粒度信息与全局上下文信息融合，用于烟叶烘烤温度和湿度预测。实验结果表明，与表现次优的模型相比，SFGLN模型仅根据烟叶烘烤图像在2度内的温度预测准确率提升4.76%，在1度内的湿度预测准确率提升了0.49%，由此可见所提模型在烟叶烘烤温湿度预测性能更优。

关键词: 温湿度预测, 烟叶烘烤, 图像识别, 卷积神经网络, 特征融合