弱监督场景下的目标检测

WEAKLY SUPERVISED OBJECT DETECTION

吕佳辉

11849190

硕士

信息与通信工程

吕佳辉

2020-05-28

2020-07-24

哈尔滨工业大学

深圳

目标检测是目前计算机视觉领域中的重要分支，机器通过视觉对现实图像中的物体进行分类并找出物体的边界框，可广泛应用于智能监控、无人驾驶、图像检索等。近些年来随着卷积神经网络的快速发展，基于深度学习的目标检测算法取得了长远的进步。但是现有的目标检测算法通常需要使用大量带有目标物体位置信息标签的数据来进行训练，而获取这样的数据需要付出很大的代价，耗费大量的人力。为此，研究者提出了在弱监督场景下的目标检测算法。这些算法只需要带有图像级标签的数据作为监督信息进行训练，大大降低了获取数据的难度，具有很大的现实意义。它吸引了越来越多研究者的关注，逐渐成为计算机视觉领域的研究热点。尽管弱监督目标检测算法很有吸引力，但是无疑增加了算法训练难度。由于训练时缺少目标物体的位置信息，使得算法很难将图像中不同类别的目标物体识别出来。 本文主要研究在弱监督场景下如何训练高性能的目标检测器。文章调研了基于弱监督学习的相关目标检测算法。目前大多数的弱监督目标检测算法都是采用多示例学习的思想。文中将详细介绍其中经典的模型，并在此基础上探究了不同后端网络对模型实验结果的影响。最后的实验结果表明选择合适的后端网络能够有效提高模型的整体表现。同时，目前主流的算法在提取特征时都存在缺少对图像全局信息捕捉的问题。因此本文提出了基于注意力机制的目标检测算法。它能有效地捕捉了图像中的全局信息，有利于识别目标物体。同时，受到有监督学习目标检测算法的启发，我们在得到带有伪标签的训练数据后，额外引入了检测分支，来进行边界框的再训练，同时这样能够做到端到端学习。大量的实验结果表明，本文所提出的基于注意力机制并结合边界框再训练的算法能有效提高在弱监督场景下目标检测的性能。

Object detection is an important branch in the field of computer vision. It is a machine that uses vision to classify objects in real images and find the bounding box of the objects. It is widely used in intelligent monitoring, unmanned driving, image retrieval, etc. With the rapid development of convolutional neural networks in recent years, object detection algorithms based on deep learning have made long-term progress. However, the existing object detection algorithm usually requires a large amount of data with the location information of the object for training, and obtaining such data requires a great price and requires a lot of manpower.To this end, the researchers proposed an object detection algorithm in weakly supervised scenarios. In this case, data only with image-level tags are needed as supervision information, thus greatly reduces the difficulty of obtaining data. The algorithm of object detection based on weakly supervision has great practical significance. Therefore, it has attracted more and more researchers' attention and become a research hotspot in the field of computer vision. Although weakly supervised object detection is very attractive, it greatly increases the difficulty of algorithm training. The lack of position information of objects during training makes the algorithm difficult to recognize different types of targets in the image.In this research, the training of high-performance target detectors in weakly supervised scenarios is investigated. The object detection algorithms based on weakly supervised learning are reviewed. Most commonly used weakly supervised object detection models are based on multi-instance learning. A typical model of multi-instance learning is introduced in detail, and the impact of different back-end networks when applied to model is explored. The experimental results show that choosing the right back-end network has an importance to the model. Meanwhile, the current mainstream algorithms lack the ability to capture image information when extracting features. Hence, this thesis proposes an object detection algorithm based on the attention mechanism, which can effectively capture the global image information, facilitating to identify the object. Moreover, inspired by the object detection algorithm of supervised learning, after the training data with pseudo labels are obtained, the detection branch is introduced and the retraining of the bounding box to achieve end-to-end learning. Experiments show that the proposed algorithm based on the attention mechanism combined with the bounding box retraining, improves the performance of object detection in weakly supervised scenes.

目标检测 深度学习 弱监督学习

object detection deep learning weakly supervised learning

中文

联合培养

南方科技大学

吕佳辉. 弱监督场景下的目标检测[D]. 深圳. 哈尔滨工业大学,2020.