复杂场景微小人脸检测与比对是一项极具挑战且未被深入研究的课题。本项目结合对抗生成网络、卷积神经网络以及图模型卷积神经网络理论，提出具有新的表达、学习和推理的对抗生成深度模型，从多维上下文信息增强、人脸结构保持以及图像转化代价评估三个方面系统研究复杂场景微小人脸检测与比对。针对人脸检测过程中感兴趣目标卷积特征消失的问题，本课题研究基于多任务学习卷积神经网络的微小人脸检测模型；针对人脸比对过程中图像质量不对称的问题，本课题研究基于双阶段生成对抗网络模型微小人脸超分辨率复原方法；针对微小人脸超分辨率复原过程中图像复原代价评估问题，本课题研究生成对抗网络中图像转化代价评估方法。本项目具有广泛的应用前景，例如，军事国防中小目标早发现、机器人视觉中小目标准确抓取和比对以及公安刑侦系统中犯罪嫌疑人准确且快速地检索。

Tiny face detection and verification in complex scene is a challenging yet not deeply studied issue. In this project we propose a novel generative adversarial model with new representation, learning and inference which integrated generative adversarial network, convolutional neural network and graph model convolutional neural network. It systematically studies complex-scene based tiny face detection and verification in terms of information augmentation, face structure preservation and image transformation cost evaluation. To tackle the problem of target convolution feature missing in region of interest during face detection, this project proposes a tiny face detection model based on multi-task learning convolutional neural network. To address the asymmetry of image quality in face verification, this project proposes a tiny face super-resolution restoration method based on two-stage generation adversarial network model. To effectively evaluate the cost of tiny face super-resolution restoration, this project proposes a new method based on generative adversarial network for image transformation cost evaluation. Compared with other works, these issues explore new topics and is of utmost importance in both theories and applications. It can be applied to fields like military defense for real-time tiny object detection, accurately capture and verification of tiny targets in machine vision, and effective retrieval of suspects in public security systems.