Extracting features from infrared images using convolutional neural networks and transfer learning

•An infrared image dataset is setup and augmented to avoid overfitting.•The architecture of VGG-19 is redesigned according infrared feature extraction.•Transfer-learning is used to reduce the computation of neural network training.•A verification experiment is conducted on MobileNet.•The results demonstrate the outperformance of the proposed method. Infrared cameras are more useful than visible light cameras in dark and foggy conditions; therefore, infrared imaging is becoming an increasingly popular subject of research. Feature extraction is an important aspect of image processing, but traditional convolutional neural networks (CNNs) trained on visible images cannot be used with infrared images. This study presents a method for retraining the Visual Geometry Group 19-layer CNN (VGG-19) to extract features from infrared images. First, a thermal image dataset was obtained from public datasets; this was then augmented by flipping, zooming, shifting, and rotating the images. Next, the architecture of the VGG-19 CNN was redesigned, and transfer learning was used to fine-tune the trainable layers. It was shown that the transfer-learned neural network could extract more information from infrared images than the original network could. To verify the validity of this method, it was also applied to the MobileNet, and the transfer-learned MobileNet also produced better results.