Graphical user interface (GUI) significantly affects the lifetime of mobile applications. However, the diversity of mobile technologies, platforms, and devices makes the GUI compatibility testing of mobile applications a challenging task. In this paper, by leveraging computer vision techniques, we take the GUI compatibility testing display problem as a screenshot difference detection problem and propose an image-based method named Cross-Device Difference Detector (CdDiff) to help testers more efficiently and accurately improve mobile application compatibility. Specifically, CdDiff uses a pair of screenshots from two devices as inputs, employs a well-designed page understanding module to extract key screen elements from each screenshot, and then utilizes a discrepancy positioning module to detect and visualize the difference elements between the two screenshots in an easily recognizable way. In addition, we built a cross-device GUI compatibility testing dataset, which contains 725 pairs of user interface screenshots taken from Baidu Netdisk, Baidu App, and Baidu Tieba, displayed on ten popular smartphones. Experiments conducted on the built dataset show that the proposed CdDiff achieved an overall accuracy of 0.924, demonstrating its effectiveness.