Infrared Image Segmentation for Photovoltaic Panels




Infrared image segmentation is the basis of error detection for photovoltaic panels. In this work, the infrared image data are collected by infrared thermal imager from the view of unmanned aerial vehicle (UAV). A semantic segmentation neural network named Deep Res-UNet, which combines the strengths of residual learning, transfer learning, and U-Net, is proposed for infrared image segmentation. Residual units are applied in both the encoding and decoding path, which makes the whole deep network ease to train. A modified ResNet-34 with pre-trained weights is utilized to get better feature representation. In the modified ResNet-34, maxpooling layer is removed for reducing the loss in resolution, an additional conv1 stage is added to copy features for the corresponding decoding path and the skip block with dilated residual block is added to generate features with larger resolution and larger receptive field. A new loss function combining Binary Cross-Entropy (BCE) and Dice is proposed to get better results. Additionally, Conditional Random Field (CRF) is integrated into the model as a post-process. The experimental results show that the prediction results of the proposed model are 97.11% and 94.47% respectively on the two evaluation indexes of F1 and Jaccard index, which is better than FCN-8s, SegNet, U-Net and two descendants of U-Net and ResNet: ResUnet and ResNet34-Unet.


Deep Res-UNet

feature maps 

we propose the deep Res-UNet, a semantic segmentation neural network, which inherits both the strengths of both U-Net and ResNet. There are some benefits of our proposed model: 1) the residual unit will make the training easier; 2) the pre-trained ResNet-34 in the encoding path brings more powerful features.

Basic Block






Related Publications

Hao Zhang, Xianggong Hong, Shifen Zhou, and Qingcai Wang.Infrared Image Segmentation for Photovoltaic Panels Based on Res-UNet. In: Lin Z. et al. (eds) Pattern Recognition and Computer Vision. PRCV 2019. Lecture Notes in Computer Science, vol 11857. Springer, Cham.