Small Object Detection of Table Tennis Based on Deep Learning Network

Abstract

According to the characteristics of the visual system of table tennis robot and the fast movement speed of table tennis, this paper focuses on balancing the speed and accuracy of small object detection based on deep learning network. We propose a deep learning network which is more suitable for small target detection, improving the detection accuracy while ensuring the detection speed to meet the real-time match. Firstly, we constructed a dataset including 20000 table tennis images. Then, we use convolutional neural network with residual connection to extract image features and construct the feature pyramid to increase the performance of network for small object detection. The feature map with rich semantic information in the upper layer is fused with the layer with rich object position information in the lower layer. We propose a new data augmentation method. During training the table tennis target on each picture is copied three times, which greatly increases the ability of network learning the feature of small target. Experimental results show that our model proposed in this paper can improve the accuracy and speed of table tennis detection, and bring great significance to the design and application of table tennis visual system.