Name
Abstract | ||
|---|---|---|
The CAPTCHA (Completely Automated Public Turing Test to Tell Computers and Humans Apart) is a common and effective security mechanism applied by many websites and applications. CAPTCHA recognition is an important and practical problem in text recognition research. Compared with traditional methods, DCNN (Deep Convolutional Neural Network) has achieved competitive accuracy in CAPTCHA recognition recently. However, current CAPTCHA recognition researches based on DCNN usually use conventional convolution network, which causes high computation complexity and great computing resource consumption. Aiming at the problems, we propose an Accurate, Light and Efficient network for CAPTCHA recognition (ALEC) based on the encoder-decoder structure. The ALEC can greatly reduce the computation complexity and parameters while ensuring the recognition accuracy. In this paper, standard convolutions are replaced by depthwise separable convolutions to improve computational efficiency. The architecture utilizes group convolution and convolution channels reduction to build a deep narrow network, which reduces the model parameters and improves generalization performance. Additionally, effective and efficient attention modules are applied to suppress the background noise and extract valid foreground context. Experiments demonstrate that ALEC not only has higher speed with fewer parameters but also improves the accuracy of CAPTCHA recognition. In detail, the ALEC achieves about 4 times speed up over the standard ResNet-18 while reducing 97% parameters. |
Year | DOI | Venue |
|---|---|---|
2020 | 10.1007/978-3-030-57058-3_5 | DAS |
DocType | Citations | PageRank |
Conference | 0 | 0.34 |
References | Authors | |
0 | 5 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Nan Li | 1 | 0 | 0.34 |
| Qianyi Jiang | 2 | 0 | 0.68 |
| qi song | 3 | 1 | 2.37 |
| Rui Zhang | 4 | 270 | 19.86 |
| Xiaolin Wei | 5 | 7 | 8.27 |