Poker Watcher: Playing Card Detection Based on EfficientDet and Sandglass Block

Abstract

We present a neural network to detect playing cards in real poker scenes through a camera, where the playing card area represents only 0.7% of the shot table area. In the acquired images, the suits of cards are fuzzy and difficult to identify, even to the naked eye. Because of the relatively few pixels corresponding to the cards, traditional image processing and pattern recognition methods struggle to detect them. Therefore, we use deep learning methods to detect, which have shown to be easy-to-use, faster and more accurate in a broad range of computer vision applications over the years. Inspired by the sandglass block, we improved the current state-of-the-art neural network architecture for object detection, EfficientDet, to retain more features. Experiments have been conducted to evaluate the performance of our improved EfficientDet model and showed that it achieved considerable performance improvement compared with the other deep learning models.