Fingerprint detection has been applied in many security applications due to its satisfying performance. Its popularity also attract the attack from an adversary who uses artificial fingerprint made of Play-Doh or silicon to mislead the decision of the system. Recently, the liveness fingerprint detection using the textural measures based on the descriptor named Binarized Statistical Image Features (BSIF). BSIF is calculated for every pixel in a fingerprint image. However, we believe that the pixels in the middle of the fingerprint contain more information than the ones in the edge. As a result, we propose a revised liveness fingerprint detection based on BSIF with sampling from Gaussian distribution. According to the Gaussian distribution, more pixels in the middle will be sampled in comparison with the ones at the edge. The experimental results show that the proposed method is more accurate than the traditional one. It may suggests that the pixels in the middle are more informative than the ones at the edge for liveness detection.

This paper proposes a novel structure extraction approach that is able to achieve high classification accuracy and low computing burden to hyperspectral image (HSI) classification based on total variation (SETV). Specifically, a two-scale decomposition-based relative total variation (TSD-RTV) method is presented for the first time to process the information of different scales such that the structure can be well extracted. Moreover, a new weighted-average fusion method is introduced, which can reduce dimensionality and also remove the noise due to hyperspectral sensors. The support vector machine (SVM) is applied to HSI classification as a classifier. Experiments are conducted on three real hyperspectral datasets: Indian Pines, Salinas, and Kennedy Space Center. Experimental results show the outstanding performance of the proposed SETV model in terms of classification accuracy and computational efficiency.

In the study of historical Tibetan document image analysis and recognition, the complete set of characters is the basis of recognition. However, character segmentation presents certain challenges due to stroke connections between characters in the same line and across different lines, making it difficult to obtain an authentic and complete set of characters from historical Tibetan documents. To address this issue, we propose AT-CycleGAN, a method for generating characters from historical Tibetan documents that achieve zero-shot character image generation. This method leverages style transformation between Tibetan print fonts(source fonts) and historical Tibetan document fonts (target fonts) while incorporating an attention module (AM) to fuse high-level and low-level features, thereby enhancing the accuracy and style consistency of the generated images. Additionally, triplet loss is employed to further improve the generation quality of similar characters. Finally, we construct a complete category of character sets from historical Tibetan documents, providing valuable data and technical support for subsequent application-oriented research.