Application of Deep Learning for Cover Image Selection in Autohome Forum Articles

The article introduces the application of deep learning algorithms for screening cover images in Autohome forum posts, emphasizing the importance of cover images in attracting user attention in the mobile internet era.

It analyzes forum topics and establishes rules for selecting candidate images containing people or vehicles, focusing on attributes such as the number of persons or vehicles, gender and age of people, vehicle angle, and completeness of persons or vehicles.

Based on this analysis, the authors define two roles for deep learning: object localization (detecting the position of persons or vehicles) and attribute recognition (judging properties such as gender, age, and vehicle orientation). Correspondingly, they choose a two-stage detection algorithm (Faster R-CNN) for localization and Mask R-CNN for human keypoint detection to assess person completeness.

The detection pipeline uses a backbone network (VGG‑16, ResNet‑101, or MobileNet) to extract features, a Region Proposal Network (RPN) to generate candidate regions, and ROI pooling (or ROI Align) to obtain fixed‑size feature maps for classification and regression.

For attribute recognition, a lightweight MobileNetV2 model is employed to classify gender (binary), age (0‑100 years, 101 classes), and vehicle orientation (eight categories: front, rear, 45° left/right, side left/right).

Experimental results show that using ResNet‑101 as the backbone for Faster R-CNN yields accurate bounding boxes for persons and vehicles, while Mask R‑NN detects 17 human keypoints; combined with MobileNetV2 attribute classification, the overall cover‑image selection accuracy reaches 81.5% on forum data.

The method has been deployed to support automatic cover‑image selection for Autohome forum posts, reducing manual effort and resources.

The article concludes with acknowledgments, references, and a note that images in Chapter 5 are sourced from Pixabay.