Professor Jin Lianwen's Team Won the Championship in the ICDAR2021 Online Signature Certification CompetitioW
time: 2021-09-15

Figure 1. ICDAR-2021-SVC Champion Award Certificate

ICDAR 2021, the flagship international conference in the field of optical character recognition (OCR), opened on September 5 in Lausanne, Switzerland. The DLVC-Lab team composed of postgraduate student Jiang Jiajia, doctoral student Lai Songxuan, and postgraduate student Zhu Yecheng under the guidance of Professor Jin Lianwen of our school participated in the ICDAR 2021 Competition on On-Line Signature Verification(ICDAR-2021-SVC), and defeated top teams from Germany, Russia, Spain, Italy, India and other countries, and once again won the international competition championship in the field of handwriting recognition by a large margin (Fig 1).

Signature authentication is an important identity authentication technology. Its authentication object is the writer's signature or its abbreviation, and it has a strong personal style due to frequent writing.Compared with features such as faces, iris, fingerprints, and voiceprints, handwritten signatures can be collected in a non-intrusive and more user-friendly way.Therefore, signature authentication has been widely used in business activities, bank office, security authentication and other scenarios.Since entering the information age, with the popularization of electronic equipment, online handwritten signature authentication technology has been extensively developed, and the acquisition medium has evolved from the original dedicated equipment in the office scene to the current mobile terminals such as smart phones and electronic tablets.In these scenarios, writers can freely choose stylus input or finger input.However, online handwritten signatures are characterized by a small number of samples, large intra-class differences in cross-time and cross-device scenarios,and are vulnerable to counterfeit signature attacks (Fig 2), bringing great challenges to the online signature verification task.


Fig 2. Schematic diagram of the trajectory of some online handwritten signatures (from the u1010 user of the DeepSignDB dataset)

Targeted at the difficulty of the problem, Professor Jin Lianwen’s team proposed a Deep Soft-DTW (DSDTW) model that can be trained end-to-end, and endows it with a classic Dynamic Time Warping (DTW) method to represent learning ability.First, they extract the time function of the speed of the online signature sequence and its first-order derivative, acceleration, pressure, angle and other information.Secondly, each time function is sent to the convolutional recurrent neural network CRNN to further learn the deep representation, and provide effective input for DTW.Then, considering that DTW is not completely differentiable to the input, its smooth form soft-DTW is introduced, and the soft-DTW distance of the signature pair is integrated into the triple loss function for optimization.Since soft-DTW is differentiable, the entire system can be trained end-to-end, achieving an elegant fusion of deep neural networks and classic dynamic time warping algorithms.The technical solution from the team of Professor Jin Lianwen stood out among the multiple methods submitted by the participating teams from Russia, Germany, Spain, Italy, India, etc., and won the championship with a large advantage (Table 1 and 2). Their method achieves an equal error rate of 3.33% in the office scene (3.11% ahead of the second place),7.41% in the mobile scene (2.73% ahead of the second place), and 6.04% in the mixture of office and mobile scene (3.92% ahead of the second place).

Table 1. Final evaluation results of ICDAR-2021-SVC

Table 2. Final score ranking of ICDAR-2021-SVC

Note: DLVC-Lab in the above table is the abbreviation of Deep Learning and Visual Computing Laboratory of South China University of Technology.

The composition of the team of Professor Jin Lianwen (DLVC-Lab) is as follows:

 Jiang Jiajia (Postgraduate student)

 Lai Songxuan (doctoral student)

 Zhu Yecheng (Postgraduate student)

 Jin Lianwen (Instructor)