基于卷积神经网络的车辆语义分割和实例分割

【作者】 张乐；

【导师】 张志梅；

【作者基本信息】 青岛大学 ， 软件工程（专业学位）， 2019， 硕士

【摘要】 计算机视觉领域中最近几年新兴两项复杂且难度又大的任务就是对图像语义分割和实例分割。随着深度学习中神经网络算法的普及,这两项任务在实际应用中变得广泛而重要。在智能交通方面,快速准确的对交通场景中的车辆进行分割是智能交通系统中的重点研究的技术之一,所以对环境多变的交通场景中的车辆语义分割和实例分割的技术研究具有重要的现实意义。随着计算机视觉步入深度学习时代,对于图像语义分割和实例分割的各个难点正在被逐渐克服,研究者们相继提出了一系列基于卷积神经网络的语义分割方法和实例分割的方法。本文主要针对于车辆语义分割对卷积神经网络模型进行改进优化以及构建对于车辆的实例分割的网络模型等方面做了研究。主要工作如下:(1)详细分析介绍了PASCAL VOC2012数据集和CITYSCAPES数据集,针对于车辆语义分割数据集不足的问题,构建了用于语义分割的数据集,即VEHICLEDATASET数据集,扩充关于车辆的语义分割的数据集。(2)本文对用于车辆的语义分割的网络复杂、且训练速度慢的问题,优化简化全卷积神经网络,将对目标的分割任务转换为基于像素的二分类问题。使用扩充后的数据集训练优化后的卷积神经网络,对比分析了分类种类不同的需求下,对网络复杂需求程度也不同的问题。在适当的降低网络复杂程度情况下,增加了网络分割的准确率。优化方法中还包括改变了网络原本使用的随机梯度下降法,选择使用Adam优化函数对网络进行训练,提高网络的训练速度。减少了网络训练的时间成本。(3)在对车辆实例分割方面,提出一种结合语义分割网络生成的掩码图和类激活图的实例分割网络模型,实现对交通环境中的车辆实例分割。图像语义分割的网络模型具有基于分类网络模型所构建的这一特点,但语义分割的掩码图只能将目标从原图中分割出来,对于同一类的实例目标却无法做出区分。而类激活技术可以直接构建在分类网络上,使用类激活技术从卷积神经网络中提取特征映射图直至生成的热力图,可以定位出不同实例在原图上的位置,实现对图像的实例区分。这样结合掩码图信息和类激活图的信息实现对车辆的实例分割。最后通过实验验证,得到了理想的实验结果。更多还原

【Abstract】 Two new and difficult tasks in the field of computer vision in recent years have been the separation of image semantics and instance segmentation.With the popularity of neural network algorithms in deep learning,these two tasks have become widespread and important in practical applications.In the aspect of intelligent transportation,fast and accurate segmentation of vehicles in traffic scenes is one of the key research techniques in intelligent transportation systems,so it is important to study the technology of vehicle semantic segmentation and instance segmentation in environmentally variable traffic scenarios.The practical significance.As computer vision enters the era of deep learning,the difficulties of image semantic segmentation and instance segmentation are gradually being overcome.Researchers have proposed a series of methods based on convolutional neural network for semantic segmentation and instance segmentation.This paper mainly focuses on the improvement of the convolutional neural network model by vehicle semantic segmentation and the construction of the network model for the instance segmentation of vehicles.main tasks as follows:(1)The detailed analysis introduces the PASCAL VOC2012 dataset and the CITYSCAPES dataset.For the problem of insufficient vehicle semantic segmentation dataset,the dataset for semantic segmentation,namely the VEHICLEDATASETt dataset,is built to expand the dataset of the semantic segmentation of the vehicle.(2)In this paper,the problem of complex network and slow training speed for the semantic segmentation of vehicles is optimized to simplify the full convolutional neural network,and the target segmentation task is transformed into a pixel-based binary classification problem.The optimized convolutional neural network is trained using the expanded data set,and the problem of the complex complexity of the network is also compared under the different types of classification.The accuracy of network segmentation is increased with appropriate reduction of network complexity.The optimization method also includes changing the random gradient descent method originally used by the network,and choosing to use the Adam optimization function to train the network to improve the training speed of the network.Reduce the time cost of network training.(3)In the aspect of segmentation of vehicle instances,an example segmentation network model combining mask map and class activation map generated by semantic segmentation network is proposed to realize segmentation of vehicle instances in trafficenvironment.The network model of image semantic segmentation has this feature built on the classification network model,but the mask map of semantic segmentation can only segment the target from the original image,but can not distinguish the instance target of the same class.The class activation technology can be directly constructed on the classification network.Using the class activation technology to extract the feature map from the convolutional neural network to the generated heat map,the position of different instances on the original image can be located to realize the instance differentiation of the image.This combines the information of the mask map information and the class activation map to implement an instance segmentation of the vehicle.Finally,through experimental verification,the ideal experimental results were obtained.更多还原