Boosting Server Type Recognition with Multi-Task, Attention and Similarity Methods

Abstract

Periodic fixed asset inventory work in the data centre room requires server type recognition and number statistics of servers. However, most of the servers are similar in appearance, making the inventory and statistics work heavy. Therefore, this paper proposes a detection and recognition deep learning model based on server similarity. We integrate the detection and recognition tasks of the server into a unified architecture for end-to-end training. At the same time, the model supports open-set testing and use. The model mainly consists of one feature encoder and two decoders for object detection and recognition. Since there is no public dataset in this field, we use the Internet Data Centre (IDC) inspection robot to collect the dataset and conduct training and testing. The speed and accuracy are higher than other detection and recognition models. In addition, it can be achieved real-time processing on the embedded device Jetson Xavier NX. Experiments show that compared with the baseline model (YOLOX model [1]), the accuracy of our model is improved by 6% and the average precision improved by 2.9%. Moreover, it can reduce the oscillation of the learning rate and accelerate the convergence.