An approach to automatically detect, localize, and repair the UI display issues in Android apps.

Our approach is named as UIBat as it is like a bat to effectively detect, localize and repair the UI display issues. And UIBat (nocturnal like bat) can complement with conventional automated program repair (diurnal like bird) for ensuring the robustness of the Mobile GUI.

We develop the UIBat platform for automatic detecting, localizing and repairing the UI display issue. Please note that we are very sorry that the domain name(url link) and server used by our UIBat now contain personal information. For the requirements of double-blind review, we give the demo video and hide the domain name information. After the blind review, we will update GitHub with the domain name(the url link) of UIBat. Up to now, UIBat detects 129 previously undetected UI display issues from GitHub, which again indicates the prevalence of UI display issues, and the necessity for the auto-repair. UIBat can repair 121 (94%) of them. For the 121 submitted pull requests, 92 of them are successfully merged by developers, while the remaining are still pending (none of them is rejected). We provide demo video as shown in below video for helping the understanding. The video link is: https://youtu.be/qakb-Yl8FJQ

The interface example of the UIBat is as follows. UIBat can automatically detect the UI display issues of the application and automatically repair the UI display issues.

UIBat can repair 121 unique UI display issues from GitHub. We submit 121 pull requests to the develop in GitHub, 92 of them are successfully merged by developers, while the remaining are still pending (none of them is rejected). Actually, the user study also suggests the relatively high degree of satisfaction of our proposed fixing, i.e., 92 merged pull requests and 66 of them receiving “like” in Github.

When merging pull requests, the developers express thanks such as "Thank you very much for this improvement, this is great. Much appreciated!" (i.e., 4345457, PicardScanner), "Very nice! Thank you for this contribution!" (i.e., 8a76280, AlarmClock). Furthermore, developers also express their thought about UI display issues "I have to admit I didn't think of this kind of issue at first and after a bit of testing the problem seems to be a bit more complex beside these login screens." (i.e., 5fcf8e1, OpenWebWeaver).

For the idea of ``Search and Destroy'' of our UIBat, the developers of GPSlogger (a project with 1.4K stars and 1 million downloads on GitHub) give us an approving look with the message shown in below figure.

The dataset includes empirical study dataset and experimental application information. Please note that the effective evaluation dataset and useful evaluation dataset in our paper are from real-world apps on GitHub (Every app can be retrieved). In this paper, we give information such as app name, issue ID, repair commit ID, etc. Researchers can find the corresponding data (APP information, APK, source code) by retrieving relevant information. Due to space constraints, we only give the relevant information of the app and no longer give the corresponding source code and APK.

(./Dataset/) shows the dataset of UI display issue we collected in our paper, it includes Android app screenshots of the UI display issue, screenshots of issue reports from Github, issue reports of Android apps in Github, issue report and its corresponding repair commit link. Please note that one of the contributions of this paper is to provide datasets of Android UI issues. So we update the dataset regularly. Therefore, the dataset provided here are updated datasets (richer and larger than those mentioned in this article).

We apologize for the issue caused by the size limit in GitHub. We update the information on GitHub and move dataset to the appropriate platform, Zenodo. Because the dataset is too large, GitHub is limited by the file upload size. And there is a file error in the previously uploaded data. In order to facilitate the download of researchers, we uploaded the corresponding data sets to the anonymous Google Drive. The sharing link of Google Drive is: https://drive.google.com/drive/folders/1NT8RrcxBR8PGw2tdEGriCaqgsfHN5tEW?usp=sharing

The interface example of the UIBatis as follows. UIBat can automatically detect the UI display issue of the Android application and automatically repair the UI display issues.

We are very sorry. Due to the large size of the dataset, we tried to transfer multiple Online Disk(e.g., Onenote, Google Drive), and there were upload errors. In view of this situation, we have developed corresponding web pages in UIBat website for our dataset. Please note that we update our dataset in real-time on our UIBat website, and the web page is shown in the figure below. Click "learn more" of dataset on UIBat to obtain the information of this dataset page.

In order to help you better understand, we give three types of UI display issues as shown in the figure below. As shown in the below Figure, they have obvious visual differences and image features, so can be distinguished with computer vision technologies.

UIBat is named as UIBat as it is like a bat to effectively detect, localize and repair the UI display issues. And UIBat (nocturnal like bat) can complement with conventional automated program repair (diurnal like bird) for ensuring the robustness of the Mobile GUI. Automatically detect, localize and repair UI display issues based on the screenshots of the mobile apps under test.

• UIBat.rar : source code of UIBat.

Due to the space limitation of our paper, we omitted some details of our model in the detection part of the UI display issue. We will describe the model in detail here.

Given the input UI screenshot, we convert it into a certain image size with fixed width and height as w * h, and the image is normalized. Then the screenshot is input into the convolution neural network ResNet50. The Convolutional layer's parameters consist of a set of learnable filters. The purpose of the convolutional operation is to extract the different characteristics of the input (i.e., feature extraction). After the convolutional layer, the screenshots will be abstracted as the feature graph. However, with the network depth increasing, accuracy gets saturated and then degrades rapidly, it is easy to appear the vanishing / exploding gradients problem and degradation problem. Because the gradient propagates back to the previous layer, repeated multiplication may make the gradient infinitesimal. As a result, with the deeper layers of the network, its performance tends to be saturated or even rapidly decline. In order to solve this problem, ResNet50 introduces the concept of residual error to solve this problem. ResNet50 solves the degradation problem by introducing a deep residual learning framework. Instead of making each layer directly fit a desired underlying mapping, it explicitly matches these layers with a residual mapping.

After obtaining the feature map through ResNet50, we input the feature map into Region Proposal Network (RPN) module. The RPN takes the feature map (of any size) as input and outputs a set of rectangular 3 object proposals, each with an objectness score. Then, a 3x3 slide window is used to traverse the whole feature map. In the process of traversing, nine anchors are generated according to rate and scale (1:2, 1:1, 2:1) in each window center. Then, full connection is used to classify each anchor (foreground or background) and preliminary bounding boxes expression. Then the bounding box expression is used to modify the anchors to obtain accurate proposals.

According to the feature map obtained by the feature extraction module and proposal obtained by RPN module. Input it into the ROI pooling layer to calculate the proposal feature maps. Finally, the proposal feature maps are input into the classification module, and the specific category (such as component occlusion, missing image, etc.) of each proposal is calculated through the fully connected neural networks (FC) and softmax layer, and the probability vector is output. At the same time, the position offset of each proposal is obtained by using bounding box expression again, which is used to regression more accurate target detection frame.

We further assess the usefulness of UIBat in real-world practice. UIBat detects 129 previously undetected UI display issues from the 106 apps, which again indicates the prevalence of UI display issues, and the necessity for the auto-repair. UIBat can repair 121 (94%) of them. For the 121 submitted pull requests, 92 of them are successfully merged by developers, while the remaining are still pending (none of them is rejected), as shown in Table