Researchers from the Hefei Institute of Physical Sciences (HFIPS), affiliated with the Chinese Academy of Sciences (CAS), have proposed a new artificial intelligence framework for target detection that provides a new solution for fast and highly accurate online target detection in real time.

In recent years, deep learning theory has contributed to the rapid development of artificial intelligence technologies. Object detection technology based on deep learning theory is also successful in many industrial applications. Current research focuses on improving the speed and accuracy of target detection and does not take into account efficiency and accuracy. How to provide fast and accurate detection of objects has become an important problem in the field of artificial intelligence.

In this study, the researchers found that one of the main disadvantages of deep learning-based target detection technology is iterative feature extraction and aggregation of deep network structures, which leads to unnecessary computational cost.

Thus, they proposed a multi-input single-output (MiSo) target recognition framework that differs from the traditional multi-input multi-output model and reduces model complexity and inference time.

In addition, according to this framework, based on the previously proposed eRF detection theory, researchers developed three new learning mechanisms to extract hotspot information more accurately and efficiently, these are the receiver field tuning mechanism, the residual attention self-learning mechanism, and the eRF-based dynamics equilibrium sampling strategies.

“We named it M2YOLOF,” said Wang Hongqiang, who led the team, “detects objects in a single object map and works well with small objects. As fast as YOLOF (One Level Visible Only Feature), but more accurate.”

They tested it on a standard dataset and achieved an average accuracy (AP) of 39.2 at 29 fps. That’s 2.6 AP higher than the cutting-edge TridenNet-R50. This method provides a new insight for target detection research and industrial applications. Source