【作者】 陈岩；

【导师】 王洪波；

【作者基本信息】 北京邮电大学 ， 计算机科学与技术， 2020， 硕士

【摘要】 随着深度学习的不断发展和硬件计算能力不断提升,计算机视觉被越来越多的研究者所关注。目标检测作为计算机视觉中的研究热点,其主要目标是:在图像中同时给出待检测物体的分类信息和坐标信息。在实际应用中,许多其他任务是基于目标检测的结果进行的,目标检测任务效果的好坏将直接影响后续任务的表现,进而影响到最终实际应用的效果。同时更快的检测速度可以使得应用更好的满足实时性的要求。因此,对于实际应用,优化目标检测算法的检测精度和检测速度,意义重大。本文的目的就是优化目标检测算法的性能,设计更加适合于实际应用的目标检测器。一般来讲,精度和速度存在相互制约的关系,一般的优化很难做到二者同时提升。本文通过调研发现,很多实际应用场景具有待检测类别较少并且背景相对固定的特点。因此本文针对此类特定场景,同时优化目标检测算法的检测精度和检测速度,主要工作和贡献如下:(1)根据本文所限定的场景特点,对目标检测模型的卷积核进行裁剪,保留网络深度的同时缩减了网络宽度。实现了检测速度成倍的提升,同时精度下降在可接受的范围内。(2)提出了自适应Focal Loss损失函数,省去手动设置超参数的步骤,让网络可以根据训练过程中收敛的情况,自适应的调节超参数,使得网络训练结果更优,提升模型检测精度。(3)针对应用场景对检测准确率和召回率的不同要求,改进了原有的基于贪心的非极大值抑制算法,提出了两种改进后的算法:Precise-NMS和Better-NMS,分别适合于高准确率和高召回率的场景。最终在加油站、咖啡厅和自动驾驶三个场景的数据集下进行实验验证,目标检测模型速度提升5.5倍,达到了192FPS,mAP分别提升了 2.47、2.34 和 0.31。更多还原

【Abstract】 With the continuous development of deep learning and the continuous improvement of hardware computing capabilities,computer vision has attracted more and more researchers’ attention.Computer vision allows machines to have the capabilities of human vision and the ability to perceive images.Object detection is a research hotspot in computer vision.Its main goal is to provide classification information and coordinate information of the object to be detected at the same time in the image.In practical applications,many other tasks are performed based on the results of object detection.Therefore,the performance of the object detection task will directly affect the performance of subsequent tasks,and then the effect of the final practical application.At the same time,faster detection speed can make applications better meet real-time requirements.Therefore,the optimization of the accuracy performance and time performance of the object detector is very meaningful for practical applications.The purpose of this paper is to optimize the performance of the object detection algorithm and design an object detector that is more suitable for practical applications.Generally speaking,there is a mutual constraint relationship between accuracy and speed,and it is difficult for general optimization to improve both.Through investigation,we found that many practical application scenarios have the characteristics of less categories to be detected and relatively fixed background.Therefore,this paper aims at such specific scenarios,optimizing the detection accuracy and detection speed of the object detection algorithm at the same time.The main contributions are as follows:(1)According to the characteristics of the scene defined in this paper,the convolution kernel of the object detection model is trimmed to reduce the network width while retaining the network depth.The detection speed has been doubled,and the accuracy has fallen within an acceptable range.(2)An Adaptive Focal Loss function is proposed,eliminating the need to manually set the hyperparameters,so that the network can adaptively adjust the hyperparameters based on the convergence during the training process,making the network training results better and improving the model detection accuracy.(3)According to the different characteristics of different application scenarios and different requirements for accuracy and recall,the original Greedy-NMS based non-maximum suppression algorithm was improved,and two improved algorithms were proposed:Precise-NMS and Better-NMS,which is more suitable for scenarios with high accuracy and high recall respectively.Finally,experiments were performed on the data sets of the gas station,coffee shop,and autonomous driving.The object detection model speed increased by 5.5 times to 192 FPS,and mAP increased by 2.47,2.34,and 0.3 1,respectively.更多还原