Recently, astronomy has witnessed great advancements in detectors and telescopes. Imaging data collected by these instruments are organized into very large datasets that form data-oriented astronomy. The imaging data contain many radio galaxies (RGs) that are interesting to astronomers. However, considering that the scale of astronomical databases in the information age is extremely large, a manual search of these galaxies is impractical given the need for manual labor. Therefore, the ability to detect specific types of galaxies largely depends on computer algorithms. Applying machine learning algorithms on large astronomical data sets can more effectively detect galaxies using photometric images. Astronomers are motivated to develop tools that can automatically analyze massive imaging data, including developing an automatic morphological detection of specified radio sources. Galaxy Zoo projects have generated great interest in visually classifying galaxy samples using CNNs. Banfield studied radio morphologies and host galaxies derived from visual inspection in the Radio Galaxy Zoo project. However, there are relatively more studies on galaxy classification, while there are fewer studies on galaxy detection. We develop a galaxy detection model, which realizes the location and classification of Fanaroff–Riley class I (FR I) and Fanaroff–Riley class II (FR II) galaxies. The field of target detection has also developed rapidly since the convolutional neural network was proposed. You Only Look Once: Unified, Real-Time Object Detection (YOLO) is a neural-network-based target detection model proposed by Redmon et al. We made several improvements to the detection effect of dense galaxies based on the original YOLOv5, mainly including the following. (1) We use Varifocal loss, whose function is to weigh positive and negative samples asymmetrically and highlight the main sample of positive samples in the training phase. (2) Our neural network model adds an attention mechanism for the convolution kernel so that the feature extraction network can adjust the size of the receptive field dynamically in deep convolutional neural networks. In this way, our model has good adaptability and effect for identifying galaxies of different sizes on the picture. (3) We use empirical practices suitable for small target detection, such as image segmentation and reducing the stride of the convolutional layers. Apart from the three major contributions and novel points of the model, the thesis also included different data sources, i.e., radio images and optical images, aiming at better classification performance and more accurate positioning. We used optical image data from SDSS, radio image data from FIRST, and label data from FR Is and FR IIs catalogs to create a data set of FR Is and FR IIs. Subsequently, we used the data set to train our improved YOLOv5 model and finally realize the automatic classification and detection of FR Is and FR IIs. Experimental results prove that our improved method achieves better performance. of our model reaches 82.3%, and the location (Ra and Dec) of the galaxies can be identified more accurately. Our model has great astronomical significance. For example, it can help astronomers find FR I and FR II galaxies to build a larger-scale galaxy catalog. Our detection method can also be extended to other types of RGs. Thus, astronomers can locate the specific type of galaxies in a considerably shorter time and with minimum human intervention, or it can be combined with other observation data (spectrum and redshift) to explore other properties of the galaxies.

中文翻译：

---

基于改进YOLOv5的FR I和FR II射电星系形态检测研究

最近，天文学见证了探测器和望远镜的巨大进步。这些仪器收集的成像数据被组织成非常大的数据集，形成面向数据的天文学。成像数据包含许多天文学家感兴趣的射电星系 (RG)。但是，考虑到信息时代天文数据库的规模非常大，人工搜索这些星系是不切实际的，需要人工。因此，检测特定类型星系的能力在很大程度上取决于计算机算法。在大型天文数据集上应用机器学习算法可以更有效地使用光度图像检测星系。天文学家有动力开发可以自动分析大量成像数据的工具，包括开发特定无线电源的自动形态检测。Galaxy Zoo 项目引起了人们对使用 CNN 对星系样本进行视觉分类的极大兴趣。Banfield 在 Radio Galaxy Zoo 项目中研究了通过目视检查得出的射电形态和宿主星系。但对星系分类的研究相对较多，而对星系探测的研究较少。我们开发了一个星系探测模型，实现了法纳罗夫-莱利 I 类 (FR I) 和法纳罗夫-莱利 II (FR II) 星系的定位和分类。自从提出卷积神经网络以来，目标检测领域也得到了快速发展。You Only Look Once: Unified, Real-Time Object Detection (YOLO) 是 Redmon 等人提出的一种基于神经网络的目标检测模型。我们在原始YOLOv5的基础上对密集星系的检测效果做了几处改进，主要包括以下几点。(1)我们使用了Varifocal loss，它的作用是对正负样本进行非对称加权，在训练阶段突出正样本的主要样本。(2) 我们的神经网络模型为卷积核加入了注意力机制，使得特征提取网络可以在深度卷积神经网络中动态调整感受野的大小。这样，我们的模型对于识别图片上不同大小的星系具有很好的适应性和效果。(3) 我们使用适合小目标检测的经验做法，例如图像分割和减少卷积层的步幅。除了该模型的三大贡献和新点外，论文还包括不同的数据源，即无线电图像和光学图像，旨在更好的分类性能和更准确的定位。我们使用来自 SDSS 的光学图像数据、来自 FIRST 的无线电图像数据以及来自 FR Is 和 FR IIs 目录的标签数据来创建 FR Is 和 FR IIs 的数据集。随后，我们利用该数据集训练我们改进后的YOLOv5模型，最终实现了FR Is和FR IIs的自动分类和检测。实验结果证明我们改进的方法取得了更好的性能。我们的模型达到了 82.3%，可以更准确地识别星系的位置（Ra 和 Dec）。我们的模型具有巨大的天文意义。例如，它可以帮助天文学家寻找 FR I 和 FR II 星系，以构建更大规模的星系目录。我们的检测方法也可以扩展到其他类型的 RG。因此，天文学家可以在相当短的时间内以最少的人为干预定位特定类型的星系，或者可以将其与其他观测数据（光谱和红移）结合以探索星系的其他特性。