Lightweight Face Detection Method For Unconstrained Scenes

Face detection is one of the classic tasks in the field of computer vision,aiming to identify all faces in a given image and return a bounding box for each face.Face detection technology is widely used in a variety of real-life scenarios such as face payment,security access control and mobile phone photography.Currently,while the development of face detection technology for fixed scenes is relatively mature,there are still many challenges for efficient and high accuracy face detection in open environments,such as lighting,occlusion,pose,scale,etc.On the other hand,there is still much room for improvement in the speed of current face detection techniques on mobile devices.Therefore,face detection techniques still have great research value in academic research and engineering applications.Early face detection algorithms commonly used a sliding window format to extract handcrafted features,which were then fed into a classifier to detect possible face regions.This approach is only effective in detecting faces in fixed scenes and is less effective in detecting faces in open environments.In addition,this face detection method cannot be trained end-toend and the detection process is more complex.The rapid development of deep convolutional neural networks has brought great breakthroughs in the fields of object detection,image classification and semantic segmentation.Convolutional neural networks are able to extract more robust features from images and achieve superior performance in various face detection scenarios.Even though this method has a major advantage over traditional face detection methods,there is still much room for improvement in various complex face detection scenarios.Face detection in real-world scenarios is more focused on efficient inference speed,so it is impractical to achieve high accuracy by simply stacking a large number of neuron parameters.In this paper,a series of methods are proposed to achieve an efficient and highly accurate face detection network.The specific research covers the following three areas.(1)A lightweight face detection algorithm with multi-scale feature fusion is constructed.The method abandons the traditional multi-level detection structure,which is considered a core component in object detection networks,and instead proposes a network structure that efficiently combines the multi-level features generated by the backbone network.This approach enables the detection of faces in a single image using only one inference on the combined single-level feature map in the detection head,significantly reducing inference time.Secondly,the design of the detection head is improved.The detection head of the face detection network is multi-tasking,and there are both correlations and differences between these tasks.Here,a multi-task sensitive detection head is designed to effectively capture both the correlation and difference between these tasks,which can further improve the detection precision without increasing the network computing load.(2)A real-time face detector based on adaptive receptive field is proposed.As the scale of faces in images varies greatly,an adaptive receptive field module is proposed to address this problem,which is embedded in the backbone network so that the multi-level features output from the backbone network have the best receptive field for detecting faces of different scales.Next,a more lightweight feature pyramid fusion module is designed to improve the feature extraction capability of the network.Finally a dynamic label assignment method is proposed that significantly improves the accuracy of the detector and does not increase the computational consumption of the inference stage.(3)An improved YOLOv7 real-time face detector is introduced.Using a Mobile One backbone network that takes full advantage of reparameterization,a multi-branch structure is used in the training phase,which is equivalently merged into a single-branch structure in the inference phase.In addition,a novel downsampling method is designed to reduce the loss of detailed features during downsampling and improve the detection accuracy of small-scale faces,in order to address the challenge of small-scale face detection with few effective features.Finally,this improved YOLOv7 face detector is deployed to mobile phones using the Onnx model transformation tool and the NCNN mobile inference framework,and its operation is tested in practice.