面向迷彩伪装目标表示与检测的生成对抗式深度学习方法研究

Caused by comprehensive impact between camouflage texture and complex scenes, the degradation of the structured information of the image of the camouflage object poses a great challenge to the shallow detection model via bottom features, such as edges and contrast. People have an integral comprehension and cognition of the object regions on the image, which are abstracted as the semantic information. Therefore, they can identify camouflaged object effectively. From the perspective of semantic information, we introduce the latest theoretical achievements, such as deep learning and generative adversarial learning, to study the representation and detection of camouflage objects. Aiming at the weakly structured properties of camouflage images, the key technologies of deep convolution network and generative adversarial network are exploited deeply, including the topology structure, regularization method and loss measurement. Furthermore, a collaborative learning framework for deep convolution networks and generating adversarial networks is constructed. The representation and detection model for camouflage images is established. This project contributes to comprehend the mechanism of visual camouflage, solve the challenge of camouflage object detection. Meanwhile, it provides a new technical way for the design, generation and evaluation of camouflage. Therefore, it has important theoretical significance and application value.

在伪装纹理和复杂背景的共同作用下，伪装目标图像的结构化信息发生严重退化，给基于边缘、对比度等底层特征的浅层检测模型造成极大挑战。人对图像上的目标区域有整体上的理解和认知，即抽象为语义信息，因此能高效检测出目标。本课题从语义信息的角度出发，引入深度学习、生成对抗式学习等最新理论成果，研究迷彩伪装目标的表示与检测问题。针对伪装图像的弱结构化属性，深入研究深度卷积网络和生成对抗网络的拓扑结构、正则化方法和损失度量等关键技术，构建深度卷积网络与生成对抗网络的协同学习框架，建立适用于伪装图像的表示和检测模型。项目有助于深入理解视觉伪装的机理，解决伪装目标检测难题，并为迷彩伪装的设计、生成、评估提供新的技术途径，有重要的理论意义和应用价值。

为解决迷彩伪装目标的表示与检测问题，本课题研究了深度学习、生成对抗学习、自知识蒸馏学习等机器学习方法。将空间注意力、通道注意力等表示结构引入基于深度卷积网络架构的目标检测模型，增强迷彩纹理的特征表示能力；提出了基于定位置信度的矩形框排序算法，有效提升了矩形框级迷彩伪装目标检测性能。在HRNet和Resnet基础上，提出了基于金字塔知识表示与迁移的自蒸馏表示结构，构建了基于结构化相似性、KL散度、交叉墒融合的蒸馏学习度量空间，在不增加模型资源开销的前提下提升了对迷彩伪装目标的像素级检测能力。提出了基于风格迁移的迷彩纹理表示方法，并利用条件生成对抗网络将其与对抗样本技术融合，生成了能有效对抗目标检测模型的迷彩风格对抗样本。提出了基于场景颜色聚类与形状模板迭代的迷彩纹理表示方法，并将其作为约束条件引导对抗贴片的训练过程，生成了能同时对抗目标检测模型与人眼视觉系统的迷彩对抗贴片。