生物视觉和认知心理学启发的目标检测算法研究

The traditional computer vision algorithms perform poorly at detecting objects in complex dynamically changing natural scenes, and the performace of their speed, detection rate and robustness can not meet the application needs. On the contrary, biological visual system performs perfect. So the mechanisms of biological vision can provide many new clues for improving the performace of the object detection algorithms. The main contents of the study are as follows: Firstly, a coarse but fast object positioning algorithm is proposed inspired by the attention mechanisms of human visual system. Secondly, a low-level feature inspired by biological vision and a middle-level feature inspired by Gestalt theroy are designed to improve the performance of “detection rate- false positive rate” and the robustness. Thirdly, a moving object detection algorithm is proposed by mimicking the dorsal pathway of human visual system. Forthly, the special problem of detecting small object, whose length is less than 30 pixels, is studied. Inspired by the contrast mechanisms and the eye moment mechanism of human visual system, an algorithm of detecting small object in images is proposed. Besides, an algorithm of detecting small moving object in video is proposed by introducing the lateral inhibition mechanisms and the motion detection mechanisms of compound eye visual system. This project is an interdisciplinary research among computer vision, biological vision and Psychology. On the one hand it has practical significance and application value to solve the key problems of automatic target acquisition and complex environment perception,On the other hand, it may play a positive role to the progress of biological vision and Psychology.

常规计算机视觉技术难以在复杂、动态、多变的自然场景中可靠检测目标，存在速度慢、检测率低、适应性差等问题。而生物视觉的物体检测能力远胜于计算机视觉，其机理可为提高物体检测性能提供新的线索:（1）通过引入视觉注意机制设计了快速目标粗定位算法，提高目标检测速度；（2）设计具有生物视觉特性的底层特征和符合格式塔理论的中层特征，提高目标检测的“检测率-误检率”性能和适应能力；（3）模拟“背侧”视觉通路设计视频中运动目标的检测算法；（4）研究高度不超过30像素的小目标的特殊检测问题，通过引入人眼对比机制和眼动机制，设计图像中小目标的检测算法，通过引入复眼视觉的侧抑制及运动敏感机制，设计了视频中运动小目标的检测算法。本课题是计算机视觉、生物视觉以及认知心理学的交叉研究，一方面对解决自动目标获取、复杂环境感知等关键问题具有现实意义和应用价值，另一方面还有助于推动生物视觉和认知心理学的研究。

生物视觉和认知心理学的机制或模型能够为计算机视觉领域的目标检测和跟踪算法研究提供启发。本项目完成了基于生物视觉的目标检测和跟踪技术研究，从生物视觉、认知心理学和行为学的角度出发，对人类视觉系统的对比机制、视觉注意机制和眼动机制这三种机制进行了模拟，并将这三种机制有机的结合起来，提出并实现了一种受视觉系统启发的目标（包括小目标）检测和跟踪算法，并构建了基于视景的仿真实验系统，通过大量实验验证了算法的准确性、适应性和鲁棒性。另外，在深度学习成为研究热点的背景下，本项目在提取格式塔中层图像特征研究方面，探索了采用深度学习框架通过监督学习来模拟Gestalt心理学中的接近律、相似律、连续性定律、求简律和闭合律，提取符合Gestalt心理学规律的特征。本课题是计算机视觉和生物视觉的交叉研究，研究成果有利于提高目标检测算法和跟踪算法的性能，包括提升速度、“检测率-误检率”性能以及对目标分辨率、部分遮挡、背景复杂、图像质量等因素的适应能力，有利用缩小算法与实际应用需求的差距。

内容获取失败，请点击重试