College of Land Resource Engineering, Kunming Universityof Science and Technology, Kunming, Yunnan, China.
Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, Yunnan, China.
PLoS One. 2022 Aug 18;17(8):e0270356. doi: 10.1371/journal.pone.0270356. eCollection 2022.
In recent years, small objects detection has received extensive attention from scholars for its important value in application. Some effective methods for small objects detection have been proposed. However, the data collected in real scenes are often foggy images, so the models trained with these methods are difficult to extract discriminative object features from such images. In addition, the existing small objects detection algorithms ignore the texture information and high-level semantic information of tiny objects, which limits the improvement of detection performance. Aiming at the above problems, this paper proposes a texture and semantic integrated small objects detection in foggy scenes. The algorithm focuses on extracting discriminative features unaffected by the environment, and obtaining texture information and high-level semantic information of small objects. Specifically, considering the adverse impact of foggy images on recognition performance, a knowledge guidance module is designed, and the discriminative features extracted from clear images by the model are used to guide the network to learn foggy images. Second, the features of high-resolution images and low-resolution images are extracted, and the adversarial learning method is adopted to train the model to give the network the ability to obtain the texture information of tiny objects from low-resolution images. Finally, an attention mechanism is constructed between feature maps of the same scale and different scales to further enrich the high-level semantic information of small objects. A large number of experiments have been conducted on data sets such as "Cityscape to Foggy" and "CoCo". The mean prediction accuracy (mAP) has reached 46.2% on "Cityscape to Fogg", and 33.3% on "CoCo", which fully proves the effectiveness and superiority of the proposed method.
近年来，小目标检测因其在应用中的重要价值而受到学者们的广泛关注。已经提出了一些用于小目标检测的有效方法。然而，实际场景中采集的数据通常是雾天图像，因此使用这些方法训练的模型很难从这些图像中提取出具有区分性的目标特征。此外，现有的小目标检测算法忽略了微小目标的纹理信息和高层语义信息，这限制了检测性能的提高。针对上述问题，本文提出了一种雾天场景下的纹理与语义集成的小目标检测方法。该算法专注于提取不受环境影响的有区分性特征，并获取小目标的纹理信息和高层语义信息。具体来说，考虑到雾天图像对识别性能的不利影响，设计了一个知识引导模块，利用模型从清晰图像中提取的判别特征来引导网络学习雾天图像。其次，提取高分辨率图像和低分辨率图像的特征，采用对抗学习方法训练模型，使网络具有从低分辨率图像中获取微小物体纹理信息的能力。最后，在同尺度和不同尺度的特征图之间构建注意力机制，进一步丰富小目标的高层语义信息。在“Cityscape to Foggy”和“CoCo”等数据集上进行了大量实验。在“Cityscape to Fogg”上的平均预测准确率（mAP）达到 46.2%，在“CoCo”上达到 33.3%，充分证明了所提出方法的有效性和优越性。
