基于并行注意力机制的地面红外目标检测方法（特邀）

Xiaofeng, 赵晓枫 Zhao; Yebin, 徐叶斌 Xu; Fei, 吴飞 Wu; Jiahui, 牛家辉 Niu; Wei, 蔡伟 Cai; Zhili, 张志利 Zhang

doi:10.3788/irla20210290

Cited by 2 publications

(2 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Secondly, we add an 'correlation layer' 15 to the fusion of two branches for feature matching. In addition, we add attention mechanism [16][17] in the middle of encoding and decoding to enhance the feature extraction and processing capability of the network. In unsupervised training, how to construct a loss function based on the predicted displacement data and the actual physical constraint relationship between the reference and target speckle image is a key issue in unsupervised learning.…”

Section: Methodsmentioning

confidence: 99%

“…In this section, we conducted experiments to validate the effectiveness of our proposed DICNet and unsupervised neural network for displacement field measurement. We used an open-source dataset 17 The sample size in the dataset is 480 pixels. Since the maximum displacement of samples in this dataset is 16 pixels, in order to reduce the error caused by missing image edge displacement information, the experimental data below are all obtained after the sample is cropped to 464×464 pixels.…”

Section: Dataset and Network Training Modesmentioning

confidence: 99%

See 1 more Smart Citation

Digital image-correlation for 2D displacement measurement based on unsupervised neural network

Wang,

Zhou,

et al. 2023

Third International Computing Imaging Conference (CITA 2023)

View full text Add to dashboard Cite

Digital Image Correlation (DIC) is a non-contact measurement technique for deformation with a long-studied challenge to find a balance between calculation efficiency and seed point quantity. Deep learning offers a new solution to improve DIC efficiency, and supervised learning DIC methods require high-quality training data, leading to challenges in ground-truth generation that can be time-consuming. We propose a DIC method for 2D displacement measurement based on unsupervised Convolutional Neural Network (CNN) to address the problem. A speckle image warp model is used to transform the target speckle image to the predicted reference speckle image according to the predicted 2D displacement map. The predicted and original reference speckle images are compared to achieve unsupervised training. Our proposed method eliminates the need for extensive training data annotation. We conducted several experiments to demonstrate its validity and robustness. The MAE and RMSE by unsupervised learning are only 0.0681 pixels and 0.0886 pixels, respectively, demonstrating the potential of our method to achieve accuracy that is comparable to supervised methods.

show abstract

Section: Methodsmentioning

confidence: 99%

Section: Dataset and Network Training Modesmentioning

confidence: 99%