利用具备相关性和互补性的辅助视觉任务来构建多任务深度网络模型，是提升机器视觉感知能力的突破方向。针对小目标特征粗糙和信息比例小等问题，构建面向小目标检测的多任务协同深度网络，解决若干关键问题：（1）构建基于多尺度特征重建的多任务网络模型，解决图像超分辨率网络与目标检测网络的多尺度特征融合结构差异问题，提出参数共享和对抗学习的skip-connection多尺度特征重建结构；（2）构建基于上下文特征融合的多任务网络模型，提出基于语义分割和目标检测的多任务深度网络及其特征交互共享结构，研究基于注意力机制的上下文特征融合方法，并进行小目标检测的语义特征可视化解析；（3）研究基于coarse-to-fine和弱监督学习的多任务网络学习方法，解决多任务训练数据集不匹配问题；（4）构建多任务深度网络动态协同训练的马尔科夫决策过程模型，研究基于强化学习的多任务网络动态协同优化方法，提高学习效率和性能。

Constructing multi-task deep neural networks with relevant and complementary visual tasks is a breakthrough research direction to improve visual perception capability of computers. Focusing on the issues such as rough features and small percentage of information in small object detection, this project will study multi-task collaborative deep networks, and solve several key problems: (1) We will construct a multi-task network model with multi-scale feature reconstruction, solving the problem of the difference between the multi-scale feature fusion structures of image super-resolution network and object detection network. Thereby, we will propose a skip-connection network structure with multi-scale feature reconstruction based on parameter sharing and adversarial learning. (2) We will construct a multi-task network model with context feature fusion by combining semantic segmentation and object detection, where the features are interacted and shared between tasks. Also we will study a context feature fusion method based on attention mechanism, and will analyze the semantic features for small object detection by visualizing them. (3) We will study the learning methods for multi-task network based on coarse-to-fine strategy and weakly supervised learning, solving the mismatch problem of training datasets. (4) We will construct a Markov decision process model for the dynamic and collaborative training of multi-task deep network, by which we will propose an optimization method based on reinforcement learning to improve learning efficiency and performance of the multi-task network.