Multi-face detection in crowd based on improved Faster R-CNN

In order to detect multi-faces in crowds for extreme high frame rate scenario, especially those with different angle, skin tones/brightness, and partially occluded faces, an image dataset was established by using cameras within main hall of a high-speed railway train station, and an optimized face detection algorithm based on Faster R-CNN was used to recognize faces in crowd in the dataset. Firstly, using VGG16 with channel attention mechanism as the backbone network of Faster R-CNN; subsequently, integrating multi-scale feature fusion technology into the network; finally, comparing the optimized Faster R-CNN with the classic version, and with YOLOv4 and SSD which were used as benchmarks. The test results show that the detection AP is increased by 3.77% when the IoU threshold is 0.7, and is increased by 0.19% when the IoU threshold is 0.5, compared to classic Faster R-CNN without optimization; and again, the detection AP is increased by 3.41% and 23.41%, respectively, when the IoU threshold is 0.7, compared to YOLOv4 and SSD algorithms.