【作者】 刘云鹏；

【导师】 王建中；

【作者基本信息】 东北师范大学 ， 软件工程， 2021， 硕士

【摘要】 目标检测作为计算机视觉的一个基础问题,随着深度学习的进步,越来越受到学者们广泛地关注,并且取得了一系列的突破。对基于无锚点的目标检测算法也逐渐进入研究人员的视线,近两年被提出的一些无锚点算法其检测效果已赶超了部分基于锚点的目标检测算法。虽然无锚点的目标检测器性能较好,但仍然存在一些问题,如:因较依赖关键点而更受背景纹理信息的影响、更多的关注角点信息而导致最后漏检。为了解决上述的问题,本文在CenterNet算法的基础上进行了两方面的改进,分别是多关键点信息融合和加入注意力机制,提出了新的检测算法EMCenterNet(Expectation-Maximization CenterNet)。本文的主要研究内容和成果如下:1.在CenterNet算法中,检测框是通过网络的一个分支检测左上角和右下角两个关键点得到的,然后,由网络的另一个分支检测到的中心关键点来判断是否保留得到的检测框。然而,由于左上或右下角关键点的缺失,CenterNet的检测结果中会出现能准确找到检测框的中心点,但是未能得到最终检测框的情况。针对这一问题,本文提出了多关键点信息融合的方法,在其检测中心关键点的网络分支上通过检测到的中心关键点来预测一个检测框,并将其与另一网络分支得到的检测框进行融合,从而得到更全面的检测框结果。2.针对CenterNet算法中因依赖关键点而更受背景纹理信息影响这一问题,本文提出将注意力机制融入到CenterNet算法中,通过注意力机制对特征图提取局部最优解,弱化背景信息,强化前景信息,使检测左上角、右上角的网络分支能更准确地找到这两个关键点,从而提高最终的检测结果。3.我们将改进后的检测算法在MSCOCO这个目标检测公开数据集上进行了相关实验,在多关键点信息融合的实验中,取得了41.8%的准确率,较基准算法提高了0.2%;在注意力机制的实验中,取得了42.2%的准确率,较基准算法提高了0.6%;将两种方法结合在一起的实验中,取得了42.5%的准确率,较基准算法提高了0.9%,综上,可以说明本文的改进在提高检测结果准确率上是有效果的。更多还原

【Abstract】 Object detection is one of the basic problems of computer vision.Through the advancement of deep learning,it has received widespread attention and obtained a series of breakthroughs.An anchor-free object detector has been proposed in the past two years,and its detection effect has surpassed that of an anchor-based object detector.Although the performance of an anchorless object detector is better,there are still some problems,such as being more affected by the background texture information because it relies more on key points,and paying more attention to the corner information,which leads to the final missed detection.In order to solve the above-mentioned problems,this paper made two improvements on the basis of the CenterNet algorithm,namely multi-key point information fusion and attention mechanism,and proposed a new detection algorithm EMCenterNet(Expectation-Maximization CenterNet).The main research contents and results of this paper are as follows:1.In the CenterNet algorithm,the detection box is obtained according to a branch of the network to detect two key points in the upper left corner and the lower right corner.Then,the central key point detected by another branch of the network is used to determine whether to retain the obtained detection frame.However,due to the absence of key points in the upper left or lower right corner,the detection results of CenterNet will show that the center point of the detection box can be accurately found,but the final detection box cannot be obtained.To solve this problem,a method of multi-key points information fusion is proposed in this paper,in which a detection box is predicted by the detected central key points on the network branch of the detection central key points,and the detection box is fused with the detection box obtained from another network branch,so as to obtain a more comprehensive result of the detection box.2.For CenterNet algorithm due to rely on key points the more affected by the background texture information of this problem,this paper puts forward the attention mechanism into CenterNet algorithm,through the attention mechanism for feature extraction of local optimal solution,weakening the background information,strengthen the prospect information,make the test on the upper left corner,in the top right corner of the network branch can find more accurate these two key points,so as to improve the final test results.3.We carried out relevant experiments on the MSCOCO object detection public data set.In the experiment of multi-key point information fusion,the accuracy of the improved detection algorithm was 41.8%,which was 0.2% higher than the benchmark algorithm.In the experiment of attention mechanism,the accuracy of 42.2% is achieved,which is 0.6% higher than the benchmark algorithm.In the experiment of combining the two methods,the accuracy of 42.5% was achieved,which was 0.9% higher than that of the benchmark algorithm.In conclusion,it can be concluded that the improvement in this paper is effective in improving the accuracy of detection results.更多还原