Research On Some Problems Of CNN-Based Face Detection

In recent years,convolutional neural network has achieved great success in various computer vision tasks.As an important research direction of computer vision,the performance of face detection has also been continuously improved.There are many variations in the application scenarios,which bring great challenges to the performance of face detection,such as scale,expression,posture,occlusion,illumination,and blur,etc.In this paper,the progress of CNN-based face detection is described in detail.To further improve the detection performance,several face detection methods are proposed.The main work of this paper includes the following three aspects.1.As face scale decreased,the performance of CNN-based detectors declines sharply.To address the issue of poor performance in small-scale,we propose a novel multi-scale face detector.The innovations of this method focus on small-scale faces in four aspects: Construct a multi-branch detection architecture and choose the shallow layers with more small-scale information as detection layers;The scale-sensitive anchor design is offered and more anchors are used to match small-scale faces,which can expand the coverage of the anchor size;Add feature map fusion modules of neighbouring branches and utilize the features of adjacent large-scale to auxiliary detect hard faces with small-scale;Multi-scale training and multi-scale testing strategies are simultaneously adopted to make the proposed model robust towards various scales.2.Current anchor matching method adopts fixed threshold to categorize positive samples.However,the maximum intersection over union between extreme aspect ratio faces and sampling anchors is always lower than the positive sample threshold,which leads to sampling failure.Although the anchor compensation can alleviate the insufficient sampling,it can not guarantee the quality of the compensated samples and the overall quality of the positive samples.To solve the problem of insufficient sampling,we propose a wide aspect ratio matching strategy to collect more representative positive anchors from faces with a wide range of aspect ratio.The innovation of this method is to construct variable positive sampling threshold for extreme aspect ratio faces.On the premise of basically maintaining the whole quality of positive samples,many high-quality positive anchors related to extreme aspect ratio faces can be obtained for training.In addition,the receptive field diversity module is designed to learn robust facial features in the feature enhancement stage.3.CNN-based face detection methods follow supervised learning manner.Their detection performance depends on the annotation quality of training set.However,existing face detection datasets are manually annotated,so it is difficult to ensure the annotation quality of each face.To address the issue of inaccurate face bounding-box annotation in training set,bounding-box deep calibration is proposed.The innovation of this method is to recognize misaligned annotations and replace them with model predicted bounding-boxes.It can reduce the regression loss by training detection models with calibrated annotations.In this way,the original balance between classification and regression losses will be broken,and drive detection models focusing on further reducing classification loss.