【作者】 杨娜；

【导师】 陈后金；

【作者基本信息】 北京交通大学 ， 交通信息工程及控制， 2013， 博士

【摘要】 摘要：车辆图像分割是车辆检测系统中最基本也是最重要的环节,车辆图像分割的质量直接影响后续图像处理的精度和效率。自然光照环境下,满足准确性、实用性要求的车辆图像分割模型始终是智能交通领域研究的热点和难点问题,因此,具有人类视觉特性的图像分割模型是今后智能交通信息处理系统研究的方向。针对上述研究背景,本文在脉冲耦合神经网络(Pulse Coupled Neural Networks, PCNN)模型的基础上,从理论与应用两个方面对具有视觉神经元特性的车辆图像分割方法进行了深入研究。论文主要研究内容和创新归纳如下：(1)针对传统脉冲耦合神经网络模型车辆图像分割时,车牌区域普遍存在的过分割与欠分割问题,提出了脉冲耦合神经网络模型最优参数的选取方法。通过最大类间方差算法自适应确定PCNN分割模型初始阈值,中心神经元局部邻域灰度均方差更新连接强度系数,Hebb学习规则确定连接加权系数矩阵,实现PCNN模型参数的优化。实验表明参数的优化不仅减少了PCNN模型图像分割时迭代的次数,而且提高了PCNN模型自适应图像分割的质量。(2)针对车辆图像分割中阴影对车牌图像分割的影响,提出了一种车辆阴影消除模型。该模型将优化的脉冲耦合神经网络与阴影属性相结合,无需建立背景模型与阴影模型。通过对图像灰度通道和色度通道信息分别进行分割,实现不同信息通道的车辆与阴影的分离,并将两个通道的分割结果图像合并,产生最终的消除车辆阴影的图像。实验结果表明该模型在消除车辆阴影的同时,较好地保持了图像中车牌、车标等关键信息。(3)针对光照和车体漫反射对PCNN模型图像分割的影响,提出了一种具有神经元感受野特性的脉冲耦合神经网络(Receptive Field-Pulse Coupled Neural Networks, RF-PCNN)车辆图像分割模型。该模型通过神经元感受野函数确定PCNN模型反馈域连接矩阵的结构,使其具有方向性和尺度性,从而更好地模拟视觉细胞分割图像的功能。实验结果表明该模型提高了自然环境中车牌图像分割的质量。分割结果中字符具有较高的边界检出率,较好地解决了复杂背景车辆图像分割中车牌区域存在的欠分割与过分割问题。(4)针对车辆图像分割中车牌所占比例小、位置不固定、大小不一以及分割易受光照影响的问题,提出了一种基于视觉注意机制的脉冲耦合神经网络(Visual Attention Mechanism Pulse Coupled Neural Networks, VAMPCNN)车辆图像分割模型。该模型在RF-PCNN模型的基础上实现了多尺度空间的图像分割,使多尺度目标均具有较好的分割效果,克服了车牌所占比例小、大小不一对图像分割的影响。将视觉注意机制中的数据驱动模式和任务驱动模式相结合,通过对不同尺度空间分割结果中目标特征尺度与最佳尺度的确定,实现多尺度空间中感兴趣目标的精确定位,实验结果表明该模型具有较好的最佳尺度分割定位多目标的功能。更多还原

【Abstract】 ABSTRACT:Vehicle image segmentation is the most fundamental and important step in a vehicle detection system. The quality of the vehicle image segmentation has a direct impact on the accuracy and efficiency of the subsequent image processing. However, it is always of great importance for intelligence transportation field to investigate an accurate and practical vehicle image segmentation model under the natural light environment. The segmentation model based on human vision is the research direction of intelligent transportation system for information processing in the future.Under this background, a vehicle segmentation method with visual neuron characteristic has been studied on the basis of Pulse Coupled Neural Networks (PCNN) model from both theory and application. The main research contents and innovation contributions of this thesis are summarized as follows:(1) The vehicle image segmentation using the traditional PCNN model usually suffers from troubles of over-segmentation and under-segmentation in the license plate area. Under this circumstance, a method for choosing optimal parameters in PCNN model is proposed. The initial threshold of PCNN segmentation model is set with Otsu algorithm, the connection coefficients are updated using the mean square difference of neurons within the local area and Hebb rule is used to calculate the connection coefficient matrix. The experimental results demonstrate that the optimized PCNN model can reduce the number of iterations during image segmentation and enhance self-adaptive segmentation effect of PCNN model.(2) A vehicle shadow elimination model is proposed to reduce the interference from the shadows in the license plate segmentation.The model combine the optimized PCNN with the shadow attributes. Meanwhile, there is no need to construct the background model and the shadow model. By performing the segmentation on both gray and hue component, vehicle and shadow are separated in each information channel. The image with the removed shadow is finally obtained by merging the segmentation results of the two information channel. The experimental results show that the model not only eliminates the vehicle shadow, but also keeps more details of the license plate and the car logo. Shadow elimination rate further validate the effectiveness of the algorithm for eliminating shadows.(3) To handle the sunlight and diffuse reflection of the vehicle body, a Receptive Field-Pulse Coupled Neural Networks (RF-PCNN) model is proposed, where the feedback domain linking matrix is determined by neurons receptive field model. This new RF-PCNN model has both directivits and scales.The function of visual cells to segment an image can be simulated more efficiently. The experimental results show that the RF-PCNN model improves the effect of license plate image segmentation under natural environment. A high boundary detection rate has been achieved for the characters and the over-segmentation and under-segmentation problems have been solved in the vehicle image segmentation with complex backgrounds.(4) The segementation effects of the license plate are influenced by many factors, such as small proportion, unfixed locations, variant sizes and variant illuminations. Aiming at solving the above problems, an image segmentation method based on the Visual Attention Mechanism Pulse Coupled Neural Networks (VAMPCNN) is proposed. This model realizes multi-scale space image segmentation on the basis of the RF-PCNN model, which achieves better segmentation effect for multi-scale targets and overcomes the influence of small proportion and variant sizes of license plate on image segmentation.The model combines data-driven mode with the task-driven mode in the visual attention mechanism, Through the determination of the target’s characteristic scale and the optimal scale, it can locate the interesting targets in the multi-scale space. The experimental results show that the model has the function to position segmentation multi-targets at optimal scale.更多还原