YOLOv5-ABLN: A Small Object Detection Model for Game App Interface

Mobile phones are becoming increasingly prevalent in modern life, and software testing is a very important aspect of ensuring the quality of the software. Game Apps (Applications) on mobile devices have unique characteristics, including frequent game version updates and more complex UI (User Interface) elements than ordinary APP page elements. The identification of interface elements is the first step in automated testing of game apps. However, the existing target detection algorithm may ignore some small target elements for the recognition of game GUI (Graphical User Interface) elements, resulting in low recognition accuracy. This paper aims at the automatic testing of mobile games and proposes a new model named YOLOv5-ABLN suitable for game APP GUI element recognition. The model is based on the YOLOv5s model, and the BRA (Bi-Level Routing Attention) module is introduced in the YOLOv5s Head, and the NWD (Normalized Wasserstein Distance) is introduced in the loss function calculation to form a new measurement index. To improve the model's ability to recognize clickable elements of the game interface, a dataset of 2,400 images is created and divided into training and validation sets in an 8:2 ratio. Furthermore, a test set consisting of 298 UI interfaces from six actively popular games is collected for evaluation and experimented on YOLOv5 and the improved YOLOv5-ABLN model respectively. The results show that YOLOv5 achieves a mAP (mean Accuracy Precision) of 90.2% in the task of detecting GUI elements of game applications. The improved YOLOv5-ABLN model achieves a mAP of 91.9% in the same task, which is a 1.7% improvement over the original model.