基于YOLOX的移动应用显示异常检测研究

Research on Display Anomaly Detection of Mobile Applications Based on YOLOX

摘要: 针对现有移动应用中显示异常检测算法检测精度不高、检测性能低下的问题，提出一种基于YOLOX的目标检测模型实现对四类常见显示异常的检测工作。针对小异常目标缺检、漏检问题，扩展特征融合网络的特征层输入尺度并进行特征融合，同时添加一个针对小目标的检测头，提升对小异常目标检测能力。通过引入坐标注意力机制，将位置信息和通道信息结合来增强网络对显示异常的识别。最后通过GIoU定位损失，实现高精度检测定位。实验结果表明，改进的YOLOX算法最终mAP达到81.10%，相比较YOLOX-s模型提高了2.25%，同时改进后检测模型与其他主流模型相比拥有良好的检测精度与模型泛化能力。该模型为移动应用显示异常智能化检测奠定了基础。

Abstract: Aiming at the problems of low detection accuracy and low detection performance of existing display anomaly detection algorithms in mobile applications, an object detection model based on YOLOX was proposed to detect four common types of display anomalies. Aiming at the problem of missing detection and missing detection of small abnormal targets, the input scale of the feature layer of the feature fusion network is extended and the feature fusion is performed. At the same time, a detection head for small targets is added to improve the detection ability of small abnormal targets. By introducing the coordinate attention mechanism, the location information and channel information are combined to enhance the recognition of display anomalies. Finally, the GIoU positioning loss was used to achieve high-precision detection and positioning. Experimental results show that the final mAP of the improved YOLOX algorithm reaches 81.10%, which is 2.25% higher than that of the YoloX-S model. At the same time, the improved detection model has good detection accuracy and model generalization ability compared with other mainstream models. This model lays the foundation for intelligent detection of mobile application display anomalies.