Research On The Deep Method For Tiny Face Hallucination And Detection

In the fields of national defense and urban management,increasingly emerging analysis demands in video surveillance recently,such as face detection,verification,and identification.However,due to the resolution,environment,lighting,and target distance variations of the surveillance camera,the face image captured by the surveillance system may be low-quality,especially very low-resolution tiny faces acquried at long distances,adversely affects face detection and recognition.Therefore,it is important to study efficient face detection and hallucination methods.Super-resolution of facial images,a.k.a.face hallucination,focusing on obtaining high-resolution face images from surveillance video or single image.However,the actual performance of most previous hallucination approaches will drop dramatically when a very low-resolution tiny face is provided,due to the lack of an informative prior as a strong semantic guidance.Under the deep learning framework,this paper not only deeply research on very low-resolution tiny face hallucination method,but also proposed the scale-adaptive face detection method.The main points are as follows:1.It is known that actual performance of most previous face hallucination approaches will drop dramatically as a very low-resolution tiny face is provided.This paper proposed tfh-BEGAN based on boundary equilibrium generative adverarial networks(BEGAN).BEGAN is viewed as an implicit face priori model and targeted auto-encoding generator with residual blocks and skip connections to fully exploit the stronger ability for representing high-resolution faces.The results demonstrate that the auto-encoding generator is a key component for tfh-BEGAN achieving hallucination performance.However,due to the limitations of iterative algorithm,tfh-BEGAN fails to produce visually realistic outputs.2.It is found that insufficient realism of the hallucilation face images and the high complexity of the alternating training network of the tfh-BEGAN,the paper introduces context loss with enhanced image realism based on the anto-encoding generation model mentioned above,proposed tiny face hallucination with context loss(tfh-CL).The experimental results demostrate that the hallucinated faces have obvious advantages in the image quality evaluation standard FID,and tfh-CL produce visually pleasant and rich high-frequency details outputs.At the same time,the recognizable features in low-resolution faces are significantly enhanced,also the applicability of the imaginary model to real low-resolution faces is improved.3.Due to the serious lack of information of very low-resolution tiny face,the detection accuracy is greatly reduced.This paper further proposes a scale-adaptive face detection model guided by attention mechanism.Specifically,this paper exploit object detection network Retina Net as the basic architecture,effectively explores the potential of space and channel attention modeling CBAM(Convolutional Block Attention Module)to guide the deep network to automatically focus on important semantic features that are helpful for face detection.The experimental results on WIDER FACE demonstrate that as a single-stage face detection network,the method proposed can significantly improve the accuracy of face detection at various scales with only a small number of parameters.