Icons are widely used as a UI component in mobile applications. However, it is considerably difficult to detect icons because of their small size. Existing rapid object detection frameworks, such as YOLO and SSD, cannot perform well on small objects, such as icons, because of the multiple downsampling layers of convolutional networks and prediction grids. This paper summarizes our recent efforts toward the goal of small-scale icon detection. First, we describe the tiny icon collection (named MobileIcon), which contains 32,670 images spanning 20 different icon categories. This database allows us to systematically study icon detection in mobile apps and to establish a benchmark for icon recognition. We also explore a way to detect icons in real-time called IconYOLO, which detects icons at the native resolution of App snapshots. We introduce a small-scale feature module based on a modified passthrough operation to fuse deep semantic information with shallow detail features. Different from the original passthrough operation, our structure employs more upsamples and DBL units for feature maps fusing. The vanishing gradient issue can be addressed by residual units. Finally, icons can be detected on a single-scale grid rather than a multiscale grid in YOLOv3. Compared with YOLOv3, IconYOLO performs 9.63% better on the MobileIcon dataset. To further verify the generalization of our network, an experiment on the traditional PASCAL VOC dataset is performed.

图标被广泛用作移动应用程序中的UI组件。但是，由于图标尺寸小，因此很难检测。由于卷积网络和预测网格的多个下采样层，现有的快速对象检测框架（例如YOLO和SSD）无法在小对象（例如图标）上很好地执行。本文总结了我们最近为实现小型图标检测目标所做的努力。首先，我们描述一个微型图标集合（名为MobileIcon），其中包含32,670张图像，涵盖20个不同的图标类别。该数据库使我们能够系统地研究移动应用程序中的图标检测，并为图标识别建立基准。我们还探索了一种名为IconYOLO的实时检测图标的方法，该方法以App快照的原始分辨率检测图标。我们引入了一种基于改进的直通操作的小型特征模块，以将深层语义信息与浅层细节特征融合在一起。与原始直通操作不同，我们的结构使用更多的上采样和DBL单元进行特征图融合。消失的梯度问题可以通过残差单位解决。最后，可以在YOLOv3中的单刻度网格而不是多刻度网格上检测图标。与YOLOv3相比，IconYOLO在MobileIcon数据集上的性能提高了9.63％。为了进一步验证我们网络的通用性，对传统的PASCAL VOC数据集进行了实验。我们的结构使用更多的上采样和DBL单元进行特征图融合。消失的梯度问题可以通过残差单位解决。最后，可以在YOLOv3中的单刻度网格而不是多刻度网格上检测图标。与YOLOv3相比，IconYOLO在MobileIcon数据集上的性能提高了9.63％。为了进一步验证我们网络的通用性，对传统的PASCAL VOC数据集进行了实验。我们的结构使用更多的上采样和DBL单元进行特征图融合。消失的梯度问题可以通过残差单位解决。最后，可以在YOLOv3中的单刻度网格而不是多刻度网格上检测图标。与YOLOv3相比，IconYOLO在MobileIcon数据集上的性能提高了9.63％。为了进一步验证我们网络的通用性，对传统的PASCAL VOC数据集进行了实验。