21 October 2024 in IEEE Access
In recent years, deep learning has been effectively used to reconstruct speckled images created by the scattering media. However, when a network with previous data training is used to test new data, it usually exhibits poor performance due to significant differences in new data that differ in scattered media and data types. Retraining the network and collecting a big amount of new data are usually required to enhance performance, but this comes with a high cost and time expense. This article proposes a transfer learning approach to address this issue. We constructed an experimental setup for capturing infrared speckled images, which includes a laser module with a 4f resonant cavity structure capable of adjusting the aperture size of spatial filters. Speckled images are collected and preprocessed at different aperture sizes, then input into a convolutional neural network for training. Transfer learning is employed to study the reconstruction of speckled images between different aperture sizes. Subsequently, to recreate speckle patterns, a speckle reconstruction network named TCGAN (Transformer_CNN_GAN) is proposed. This network is based on pre-training TCGAN with an enormous quantity of data and fine-tuning its model with new data so that it can be adapted quickly with a small sample size. The experimental findings show that TCGAN may achieve good reconstruction without too much new data by means of transfer learning. Moreover, pretraining the network with more complex datasets helps enhance the network’s transfer learning capability. Therefore, conducting research on optical field imaging and restoration methods based on transfer learning holds significant theoretical and practical significance.
1 July 2024 in Journal of Electronic Imaging
When facing motion and complex environmental conditions, infrared videos captured by thermal imaging devices often suffer from blurring, leading to unclear or missing details and positional information about the targets. To improve this problem, this work proposes an improved deblurring method suitable for infrared videos based on a deep learning-based deblurring network originally designed for visible light images. This method is built upon the D2Net network by introducing a spatial and channel reconstruction convolution for feature redundancy, enhancing the network’s capability for image feature learning. In terms of the encoder-decoder module, a triple attention mechanism and fast Fourier transform are introduced to further improve the network’s deblurring performance. Through ablative experiments on infrared datasets, the results demonstrate a significant improvement in deblurring performance compared to the original D2Net. Specifically, the improved network achieved a 1.42 dB increase in peak signal-to-noise ratio and a 0.02 dB increase in structural similarity compared to the original network. In summary, this paper achieves promising results in infrared video deblurring tasks, demonstrating the effectiveness of the proposed method.