【作者】 刘琳；

【导师】 王刚；

【作者基本信息】 吉林大学 ， 软件工程（专业学位）， 2021， 硕士

【摘要】 近年来,计算机视觉在日常生活中的重要作用日益凸显。目标检测作为计算机视觉的基本工作之一,得到了普遍的应用,不仅可以对目标进行识别还可以对图片、视频等资料进行解释,可以说社会中的方方面面都可以看到目标检测的影子。目标检测的任务是解决目标的分类和定位的难题,但是由于物体的形态各式各样,且有背景、光线等的干扰,除此之外,物体之间还存在相互遮挡的现象,所以,努力提高目标检测的精度十分具有研究意义。当目标检测被应用在自动驾驶这种对安全性要求较高的领域、以及应用在医疗领域辅助医生进行医学诊断时,它的精确度也日益得到了大家的广泛关注。因此,更好地提升目标检测的精确度是大势所趋。传统的方式和深度学习的方式是解决目标检测问题的两种主要方式。对于传统方式,多数算法是通过人工发现特征的方式设计算法。尽管传统目标检测算法较为简单,但是仅能适用于目标单一、特征明显的情况下。然而,现实世界中的物体往往多种多样,而且背景变幻莫测,所以要找到众多物体的共同特征来实现目标的检测十分困难。对于深度学习方式,有两阶段目标检测算法和一阶段目标检测算法这两种方法。两阶段算法先在图像上生成候选框,然后再对候选框分类和回归,一阶段算法则是把分类以及定位问题看成一个回归问题来处理。正是由于两类算法的实现方式不同,所以它们的性能也有差异,两阶段算法精确度高,而一阶段算法速度快。本文对两阶段目标检测算法中具有代表性的更快的基于区域的卷积神经网络(Faster Region-based Convolutional Neural Network,Faster R-CNN)做出了改进,设计了一种新的架构——基于改进的Faster R-CNN目标检测算法,做出的工作如下所示:(1)为了使学习率的设置更加合理,找到更适合网络训练的学习率,本文采用鲸鱼优化算法对学习率的设置进行了优化。训练过程中分两个阶段来寻找最优的学习率,第一个阶段在区间(0.0002,0.002)上采用鲸鱼优化算法寻找高学习率的最优值,使得网络快速收敛。当训练达到指定迭代次数,训练进入第二阶段。第二个阶段在区间(0.00002,0.0002]上采用鲸鱼优化算法寻找低学习率最优值,直到训练达到最终指定迭代次数,实现训练收敛,达到更好的训练效果。(2)为了提高损失函数的整体性能,本文对损失函数进行了改进,引入了两个权重因子,分别用于调节分类损失和回归损失所占的权重。同时,为了使这两个权重因子的设置更加合理,本文使用了鲸鱼优化算法对这两个值进行了优化。为了验证本文提出的改进是否具有价值,在三个目标检测数据集上做了一系列实验,结果证明,本文的算法在目标检测领域中的尝试是有价值的。而且,本文算法跟几种经典的算法相比,本文的算法也获得了较好的效果。因此,能够证明本文的算法具有通用性和应用价值。更多还原

【Abstract】 In recent years,the important role of computer vision in daily life has become increasingly prominent.Object detection,as one of the basic work of computer vision,has been widely used.It can not only identify object but also interpret pictures,videos and other materials.It can be said that object detection can be seen in all aspects of society.Object detection is to solve the problem of object classification and positioning.However,due to the various shapes of objects,the interference of background and light,as well as the phenomenon of mutual occlusion between objects,it is of great significance to strive to enhance the accuracy of object detection.When object detection is applied in the area of autonomous driving that require high safety,as well as in the medical field to assist doctors in medical diagnosis,its accuracy has also attracted widespread attention.Therefore,it is an inevitable trend to further improve the accuracy of target detection.The traditional method and the deep learning method are two main methods to solve the problem of object detection.For traditional methods,most algorithms are designed by manual features.Although traditional object detection algorithms are relatively simple,they are only suitable for situations with single object and obvious features.However,objects in real life are various,and the background is changeable,so it is very difficult to find the common features of many objects to achieve object detection.As for deep learning methods,two-stage object detection algorithms and one-stage object detection algorithms are two main mehods.The two-stage algorithm first generates candidate boxes on the image,and then classifies and regresses the candidate boxes,while the one-stage algorithm directly transforms the classification and location of the object into a regression problem.Because the two types of algorithms are implemented in different ways,their performance is also different.The two-stage algorithm has high accuracy,while the one-stage algorithm is fast.In this paper,the representative Faster region-based Convolutional Neural Network in the two-stage object detection algorithm is improved.A new network architecture,the object detection algorithm based on improved Faster R-CNN,is proposed to make it have better object detection effect.The main improvements of this paper are as follows:(1)To make the setting of the learning rate more reasonable and find a learning rate that is more suitable for network training,this article adopts the whale optimization algorithm to optimize the learning rate.The training process is divided into two stages to find the optimal learning rate.In the first stage,the whale optimization algorithm is used to find the optimal value of the high learning rate in the interval(0.0002,0.002)and accelerate the training convergence speed.When the training reaches the specified number of iterations,the training enters the second stage.In the second stage,the whale optimization algorithm is used to find the optimal value of the low learning rate in the interval(0.00002,0.0002].Until the training reaches the final specified number of iterations,training convergence is realized and better training results is achieved.(2)To enhance the overall performance of the loss function,this paper optimizes the loss function and adds two weighting factors,which are used to adjust the weight of the classification loss and the regression loss respectively.At the same time,in order to make the setting of these two weighting factors more reasonable,this paper uses the whale optimization algorithm to optimize the two weighting factors.This paper has conducted experiments on two datasets,and the results prove that the detection accuracy of the algorithm proposed in this paper has obtained good accuracy in the field of object detection.Moreover,compared with several classic algorithms,the algorithm in this paper obtains better object detection accuracy.Therefore,it can be proved that the algorithm in this paper has universality and application value.In order to verify whether the improvement proposed in this paper is valuable,experiments are carried out on the three datasets of Pascal VOC 2007,Pascal VOC 2012 and MS COCO.The results prove that the algorithm of this paper is valuable in the field of object detection.Moreover,compared with several classical algorithms,the algorithm in this paper also achieves better results.Therefore,it can be proved that the algorithm in this paper has universality and application value.更多还原