基于弱监督学习的协同目标检测算法研究

As an important research in the field of computer vision, object detection is an important prerequisite for many high-level visual processing and analysing. In practical applications, due to the high cost of sample marking, weakly supervised learning has received extensive attention in recent years, while existing weakly supervised algorithms are less robust.Based on the above problems, a collaborative object detection algorithm based on weakly supervised learning is proposed.The main research contents are as follows:(1) A weakly supervised collaborative object detection model is constructed. This model uses weakly supervised detectors to learn label information, and achieves collaborative learning between weakly supervisied learning and strong supervised learning.(2)Based on network feature layer clipping, convolutional kernel weight cropping and weights quantization object detection model compression method, a high performance object detection model is constructed, which enhances the practicability of object detection.（3）Aiming at the practical application of aerial image, a full convolution deep residual network based on denoising autoencoder and weakly supervised cooperative object detection and network compression model are proposed to realize the refined detection from superclass to subclass.Through the high-performance multi-class weakly supervised object detector constructed above, the performance of the object detector is improved from the aspects of computational complexity and robustness, which has a profound practical significance and effect for the development in the field of autonomous vehicles, video surveillance and robotics technology.

目标检测作为计算机视觉领域的重要问题，是众多高层视觉处理和分析任务的前提基础。近年来，基于弱监督学习不依赖大量标签信息的优势，得到国内外学者的关注。本项目旨在研究基于弱监督学习的协同目标检测模型，并构建面向航拍影像的精细化多目标协同检测系统。研究内容如下：（1）构建标签信息自主学习的弱监督协同目标检测模型，定义权重自适应分类损失函数和目标位置回归损失函数对提出的模型并行优化，提高目标检测的鲁棒性。（2）构建基于网络特征层裁剪、卷积核权重裁剪以及权重量化的目标检测压缩模型，增强检测场景的实用性；（3）针对航拍图像实际应用，研究降噪自编码残差网络和弱监督协同的目标检测模型，实现从超类到子类的精细化协同目标检测。通过对弱监督协同目标检测模型的研究及应用，提高现有弱监督学习的鲁棒性和实用性，对未来自动驾驶、视频监控和机器人视觉等技术的发展带来重要意义和影响。

目标检测作为计算机视觉领域的重要问题，是众多高层视觉处理和分析任务的前提基础。近年来，基于弱监督学习不依赖大量标签信息的优势，得到国内外学者的关注。本项目研究了基于弱监督学习的协同目标检测架构，并构建了面向航拍影像的精细化多目标协同检测系统。主要研究内容如下：（1）针对深度学习标签数据成本过高的情况，构建了标签信息自主学习的弱监督协同目标检测架构。通过定义权重自适应分类损失函数和目标位置回归损失函数，在此基础上给出了并行优化结果，提高了弱监督目标检测分类准确率和泛化能力。（2）针对遥感目标检测器以及全卷积深度残差网络参数量大的问题，提出了网络模型压缩架构，解决了模型中参数冗余的问题。（3）针对航拍图像实际应用，提出了降噪自编码残差网络和弱监督协同的目标检测架构，实现了从超类到子类的精细化协同目标检测，解决了航拍图像存在噪声、几何畸变、运动模糊等问题。通过对弱监督协同目标检测模型的研究及应用，提高了现有弱监督学习的鲁棒性和实用性，对未来自动驾驶、视频监控和机器人视觉等技术的发展带来重要意义和影响。