生物视觉感知启发下的目标检测与识别技术研究

发布时间：2018-04-15 14:23

本文选题：目标检测与识别 + 视觉感知系统视觉感知双通路结构　；参考：《华中科技大学》2015年博士论文

【摘要】：人类视觉感知系统能够轻易地从外部复杂场景中感知静止或运动的目标,然而该感知过程对计算机来说并非易事。作为计算机视觉领域中的研究热点和难点之一,目标检测与识别长期以来受到了众多学者的关注,具有重要的学术研究价值。目标检测与识别是视频监控、智能交通、遥感测绘等应用领域的核心技术,具有广泛的应用前景。近年来,大量的自动目标检测与识别算法被提出,一定程度上提高了机器视觉感知系统的智能化水平。然而,机器视觉感知水平还远远落后于人类视觉的感知能力。鉴于这一现状,借鉴神经科学和认知心理学的研究成果,模拟生物视觉感知过程来建立目标检测与识别算法,是一个能极大程度提高目标检测与识别性能的研究和探索方向。在生物视觉感知特性启发下,本文开展了一系列目标检测与识别技术研究,着重研究了空间显著目标检测技术、时空显著目标检测技术、机载红外运动目标检测技术和无监督层次化特征提取及其在分类识别中的应用技术,具体内容如下： (1)基于选择性视觉注意中的感觉增强理论,提出了一种基于柯西图嵌入的目标空间显著性计算方法。感觉增强理论表明目标的注意是注意焦点扩散的结果。基于这一理论,该方法以经典注意焦点预测方法为输入,利用基于柯西图嵌入的平滑约束对输入结果进行平滑、完成视觉组织过程,用以模拟注意焦点的扩散过程。实验结果表明基于柯西图嵌入的视觉组织优于传统的拉普拉斯图嵌入,可以大大提高原输入方法的显著目标预测性能,并且最终结果达到与已有专注于显著目标检测的方法可比的性能。 (2)在视觉感知双通路结构启发下,提出了一种基于混合特征核回归模型的目标时空显著性计算方法。视觉感知双通路假设被普遍接受：视觉感知首先沿着腹部流通路和背部流通路分别进行,最终完成信息整合。在这一视觉感知流程启发下,提出的时空显著性计算方法包括三个主要模块：空域显著性计算对应腹部流通路的外观感知功能；时域显著性计算对应背部流通路的运动感知功能：时空显著性融合对应双通路的信息整合功能。为了应对时空显著性的融合问题,首次提出混合特征空间核回归模型诱导下的三种估计子,包括常数估计子、线性估计子和非线性估计子。结合三种估计子,创造性的提出了一种混合融合算法。实验结果表明所提出的混合融合算法明显优于已有时空显著性融合算法。得益于该混合融合算法,所提出的时空显著性计算方法取得比已有方法更好的性能。 (3)在背部流感知通路反馈连接启发下,提出了一种多级处理算法来解决机载红外运动目标检测问题。大量的神经生理学实验表明负责运动信息处理的背部流通路的处理速度远远快于负责外观信息处理的腹部流通路,并且呈现出明显的背部流通路反馈调节腹部流通路的特性。在这一不对称连接特性启发下,提出一种实时机载红外运动目标检测算法,主要包括三级处理流程：快速运动感知；候选运动焦点及运动窗获取；运动窗内目标精确提取。实验结果表明提出的多级处理算法在计算效率和检测精度两方面都优于已有的算法。 (4)在视皮层层次化感知特性启发下,提出了一种无监督层次化特征提取方法并应用于目标和场景分类识别任务。大量的神经生理学证据表明人类具有优越的目标感知和判别能力是因为视皮层的不同皮层区域能够响应由低级到高级不断抽象的特征。具体来说,初级视皮层V1区细胞对类Gabor滤波器的低阶边缘比较敏感,而高级视皮层V2区细胞对高阶的角点、连接点等特征比较敏感。从可视结果来看,所提出的层次化特征提取方法中的第一级特征表达的基函数类似于V1区细胞响应,并且第二级特征表达的基函数类似于V2区细胞响应。该特征提取方法在10类目标分类任务的CIFAR-10数据库和21类遥感影像场景分类任务的UCM-21数据库上,均取得了理想的分类效果。
[Abstract]:In recent years , a large number of automatic target detection and recognition algorithms have been put forward in order to improve the intelligence level of the machine vision perception system .

In the light of biovisual perception , a series of target detection and recognition techniques are studied in this paper , focusing on the research of the techniques of spatial salient object detection , space - time salient object detection , airborne infrared moving target detection and non - supervised hierarchical feature extraction and its application in classification recognition , which are as follows :

( 1 ) Based on the perception enhancement theory in selective visual attention , a method for calculating the target spatial significance based on Cauchy ' s embedding is presented . Based on this theory , the method is based on the classical focus prediction method , which is used to simulate the diffusion process of focus . The experimental results show that the embedded visual organization is superior to the traditional Laplacian image embedding , which can greatly improve the prediction performance of the original input method .

( 2 ) In the light of the visual perception double - path structure , a method for calculating the objective time - space significance based on the hybrid characteristic kernel regression model is proposed . The visual perception double - path assumption is generally accepted : the visual perception is firstly carried out along the abdomen flow path and the back flow path respectively , and finally the information integration is finished . Under the inspiration of this visual perception process , the proposed temporal and spatial significance calculation method comprises three main modules : the airspace significance calculation corresponds to the appearance perception function of the abdominal flow path ;
In order to solve the fusion problem of space - time significance , three kinds of estimation sub - parameters , including the constant estimator , the linear estimator and the non - linear estimator , are proposed for the first time in order to deal with the fusion problem of space - time significance . The results show that the proposed hybrid fusion algorithm is superior to the existing spatial - temporal significance fusion algorithm .

( 3 ) Inspired by the feedback connection of the back flow sensing path , a multi - level processing algorithm is proposed to solve the airborne infrared motion target detection problem . A large number of neurophysiology experiments show that the processing speed of the back flow path for the motion information processing is much faster than that of the abdominal flow path responsible for the appearance information processing .
candidate motion focus and motion window acquisition ;
The experimental results show that the proposed multi - level processing algorithm is superior to the existing algorithms in terms of efficiency and precision .

( 4 ) Based on the hierarchical cognitive characteristics of visual cortex , a hierarchical feature extraction method is proposed and applied to target and scene classification recognition tasks .

【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【相似文献】