基于不动点方程解析求解的高动态场景多尺度分割

Exploring multiscale modeling strategy and interscale coupling mechanism of highly dynamic scenes to achieve reliable segmentation remains to be a big challenge for image processing. The key is to couple uniformity of high-scale features with similarity of low-scale cues as well as accuracy of the algorithm. Existing methods usually lack position constraints of overall target features detected in large space of original images and the consistency between coupling posteriori with true distribution can not be guaranteed. We propose to detect targets in the lower dimensional wavelet approximation coefficients to apply overal position constraints and improve coupling accuracy via analytical solution of fixed-point equation. We plan to investigate target contour detection and position mapping methods in high-scale approximation coefficients so as to couple low-scale color and texture similarity in local features. To investigate belief propagation strategy for exact inference of posterior distribution by obtaining analytical solution of fixed-point equation of the convex dual, which is derived from variational lower bound of information projection between posterior and real distribution. Carrying out segmentation experiments on highly dynamic scene benchmarks as well as images acquired by robot vision platform and compare with SWA, MNC and gPb algorithms. The scientific objective is to explore coupling mechanism of overall and local features as well as address non-consistency between posterior inference and real distribution to achieve exact inference of highly dynamic scenes for the purpose of robust segmentation.

探索高动态场景图像多尺度建模策略与尺度间耦合机制实现可靠分割是图像处理领域的一大挑战，其关键是低尺度局部线索相似性与高尺度整体特征统一性耦合以及算法求解准确性。现有方法通常在缺乏整体位置约束且搜索空间较大的原图像域寻找目标整体特征，耦合后验与真实分布一致性也无法保证。本项目提出在维度较低的小波近似系数空间检测目标并施加整体位置约束，用不动点方程解析求解提高耦合准确性。研究高尺度近似系数内目标整体轮廓检测及位置映射，耦合低尺度色彩和纹理相似性局部特征。研究后验与真实分布信息投影变分下界凸对偶及其不动点方程解析求解并确定信度传播策略，准确推理多尺度后验分布。用高动态场景标准库和机器人视觉计算平台获取高动态图像开展分割实验，与SWA、MNC、gPb多尺度分割算法进行性能对比。本项目旨在探索整体与局部特征多尺度耦合机制，解决耦合后验与真实分布不一致难题，为动态场景下准确推理和鲁棒分割提供理论依据。

本研究针对现有分割通常在缺乏整体位置约束图像域寻找目标特征，耦合后验与真实分布一致性无法保证的局限，研究在维度较低的小波近似系数空间检测目标并施加整体位置约束，用不动点方程解析求解提高耦合准确性。其主要内容包括高尺度近似系数目标整体轮廓检测及位置映射，耦合低尺度色彩和纹理相似性局部特征；后验与真实分布信息投影变分下界凸对偶及其不动点方程解析求解和信度传播策略，准确推理多尺度后验；开展高动态场景图像分割实验，与先进分割算法对比并进行定量评价。研究过程中建立了高尺度目标假定位置检测模型，组合时域运动边缘和空域谱边缘构成高尺度整体位置区域，提取显著性引导目标帧间特征非线性回归达到前景目标时域一致排序，提出拉格朗日乘子近端投影和稀疏映射显著支撑投影，将小波域多尺度显著映射投影到马尔科夫随机场，提高了动态场景后验概率与真实分布一致性。引入权重总变分能量泛函全局无约束凸优化，将权重总变分目标函数的优化问题转化为不动点方程无约束凸优化，应用交替方向不动点方程迭代解析求解总变分能量泛函并制定了多尺度显著流消息传播策略，达到了更具辨识能力的边界特征跨帧传播和高质量的显著目标映射。采用高动态场景图像分割标准库开展分割实验，定性实验结果表明本项目提出的权重总变分不动点方程解析求解得到的显著前景映射最优解使得前景种子点映射更为有效平滑。弗莱堡-伯克利运动分割标准库30个训练集和29个测试集定量评价结果表明，本项目算法分割精度分别达到75.89%和71.22%，比快速目标分割方法分割精度提高15.26%和17.19%，全部59个动态场景视频集本项目方法分割精度达到73.56%，比无向图视频目标分割方法提高5.55%。验证了本项目提出的多尺度显著流分割算法能跨时域传播多尺度近似系数表征的显著映射，克服了大位移前景目标造成的运动边缘混淆。本项目提出的不动点方程解析求解图像分割算法，提高了分割模型后验分布一致性，为高动态场景鲁棒分割提供了理论指导和依据。