Cutting edge technology

Introduction: Recently, the face detection algorithm VIM_FD independently developed by Zhongxing Micro AI surpassed the Baidu face detection deep learning algorithm PyramidBox (0.961) with the excellent score of 0.967 in the face detection competition of WIDER FACE. Tencent's DSFD algorithm (0.966) and Huawei Cloud FDNet face algorithm (0.959) won the first place in the world.

Recently, the face detection algorithm VIM_FD independently developed by Zhongxing Micro AI surpassed the Baidu face detection deep learning algorithm PyramidBox (0.961) with a superior score of 0.967 in the face detection competition of WIDER FACE. Tencent's DSFD algorithm (0.966) and Huawei Cloud FDNet face algorithm (0.959) won the first place in the world.

WIDER FACE is one of the largest face detection data sets in the industry with the largest data size (about 400,000 face marks) and the most difficult detection. It has gradually surpassed FDDB as a research institution and company. Industry benchmark.

VIM_FD won the world first in the WIDER FACE face detection competition, indicating that the Vimicro AI has been ranked first in the world in the field of face recognition.

Zhang Yundong, chairman of Beijing Zhongxing Micro Artificial Intelligence Chip Technology Co., Ltd., said that Zhongxing Micro AI takes into account the data size of different scenarios, resulting in a single processor unable to meet all data computing needs of various scenarios. Multi-mode computing, multi-core architecture is A trend in the development of AI processors, multi-modal computing includes three scene modes: deep learning algorithms for dealing with big data scenes; traditional intelligent algorithms for dealing with small data scenes; and scenes for processing logical reasoning programs based on abstract logical thinking. In order to support multi-mode computing, the chip itself must be implemented by multiple cores: the deep learning algorithm for big data should be implemented by NPU, and the traditional intelligent algorithm for small data should be implemented by DSP, which should be adopted for the scene of processing logical reasoning program. CPU to achieve. Three core functions in one chip generate enough data analysis interactions for multiple processors to form a closed loop with tight coupling and strong correlation. To support artificial intelligence analysis and calculations. This is the development trend of integrated circuits in the post-Moore era.

It is reported that the V proposed by Zhongxing Micro AIThe IM_FD face detection algorithm adopts the STC and STR module design ideas, and combines multiple layers of features to improve the accuracy of frame positioning by combining high-level, low-resolution, strong semantics, etc., while using low-level, high-resolution, and weak Semantic information to enhance the network's ability to handle small targets. In addition, the Attention mechanism is used to provide high semantic and class-aware features to activate and calibrate feature maps for target detection. It is worth mentioning that the same algorithm model does not undergo any tuning. Under the FDDB Continuous Score evaluation index, it also leads all the competitors and university research institutions at home and abroad, which proves its excellent generalization ability.

Hot topic