Cascade CNN Detector For High-Density Face Detection

With the continuous development of computer vision technology,people's demand for image processing has gradually improved from low-level needs such as photography,recording,playing,etc.,to intelligent needs such as classification,detection,recognition,segmentation,etc.Therefore,a large number of computer vision tasks and functions have been continuously expanded and improved.As the most popular computer vision issue in recent years,face attracts a lot of related research and work,including face detection,face recognition,3D face reconstruction,face synthesis,face tracking,etc.Among them,face detection is one of the most important issues as a prerequisite of face-related applications.Although face detection and analysis technology have developed for many years,it is still a thorny problem.Since the face-related applications gradually changed from constraining methods to unconstraining methods,from the single face serial scenarios to high-density face parallel scenarios,the requirements for face detection methods are also increasing.Influenced by the background,lighting,occlusion,pose and other,in high-density scenes,the difficulty of face detection is greatly improved.In addition,with the increasing requirements to utilize the face detection method in computing-limited mobile devices requires the face detection model to be as lightweight as possible.Compared with a single complex network,the cascaded convolutional neural network can effectively solve the problem of imbalance between positive and negative samples in face detection.Most of the easy background will be rejected in the early part in cascade face detectors which shows great efficiency to satisfy the real-world requirements.However,in the high-density and unconstrained scenarios,there is still a big gap between its accuracy and that of a single complex network method.Therefore,this paper analyzes the cascaded face detection algorithm and proposes a lightweight cascaded neural network for high-density face detection which significantly improves the performance of the cascaded face detection method in high-density unconstrained scenarios.In summary,the main work of this paper includes the following three parts:(1)Due to the limited representation ability of the lightweight proposal network,there are a large number of negative samples that can not be rejected by the proposal network and enter the refine networks,which greatly decrease the speed of the whole process.Observe the above condition,we propose a proposal network with skin-aware pixel-wise attention.By embedding a skin color-sensitive attention module,we can assist the proposal network to reject easy negative samples.It significantly reduces the total number of proposal boxes under the same recall rate,so as to improve the detection efficiency of the whole cascade network.(2)Considering the correlation and difference between the two tasks of face classification and box regression in the process of face detection,if the two tasks use completely independent features,the correlation between the tasks can not help the backbone to optimize the feature extracting;if the two tasks use completely shared features,they cannot achieve the best performance of each task.For this reason,we proposed a Separated Attention Module(SAM)based on feature channel attention,which can construct attention modules for these two tasks respectively.The SAM let these two tasks share the low-level features and extract the high-level features of their respective tasks at the same time to improve their performance.(3)Since there exists a large number of small-scale faces in the high-density scenarios,these small-scale faces contain few detectable information and are easy to be confused with the background.Because lightweight cascade face detectors have weak feature representation,the cascaded network is not able to represent the small-scale face,which causes lots of false-positive.For this reason,this paper proposes a new structure module for context information enhancement,which is attached at the end of the first level proposal network and uses more context information to better connect the proposal network and the purification network,which greatly improves the performance of cascaded face detection in high-density scenarios.