Abstract
At present, most departments in colleges have their own official accounts, which have become the primary channel for announcements and news. In these official accounts, the popularity of articles is influenced by many different factors, such as the content of articles and the aesthetics of the layout. This paper mainly studies how to learn a computational model for predicting page views on college official accounts with quality-aware features extracted from pictures. First, we built a new picture database by collecting 1,000 pictures from the official accounts of nine well-known universities in Beijing. Then, we proposed a new model for predicting page views by using a selective ensemble technology to fuse three sets of quality-aware features. Experimental results show that the proposed model has achieved competitive performance against state-of-the-art relevant models on the task of inferring page views from pictures on college official accounts.
Introduction
In the digital age, university departments rely on official accounts (specifically on platforms like WeChat) as their primary news and announcement channels. While the text of an article is important, the visual appeal of the featured images often acts as the "hook" that determines whether a user clicks or scrolls past. The challenge lies in quantifying "quality." Traditional image metrics often focus on technical resolution, but human perception is far more complex. This research addresses the gap between raw image data and human engagement by asking: Can we build a model that "sees" image quality the way a student does and uses that to predict an article's popularity?
Methodology
The researchers developed a selective ensemble framework that extracts and fuses 99 distinct features across three categories:
- NSS-Based Features (36 features): Based on Natural Scene Statistics, these identify how much an image deviates from the statistical properties of a "perfect" natural scene, detecting blurs, noise, or distortions.
- Histogram-Based Features (40 features): These focus on the Human Visual System’s (HVS) sensitivity to structure. They use Gradient Magnitude (GM) to capture contours and Laplacian of Gaussian (LoG) operators to mimic how retinal cells perceive contrast.
- Free Energy-Based Features (23 features): Inspired by cognitive science, this approach assumes the brain has an internal generative model of the world. High-quality images are easier for the brain to "predict," while low-quality images create more computational "surprise" (entropy).
The team used a selective ensemble technology to combine these features. Rather than relying on a single algorithm, they fused these diverse feature sets to improve fitting accuracy and prevent overfitting.
The Official Accounts Picture Quality Database (OAPQD)
To train and test the model, the researchers introduced the Official Accounts Picture Quality Database (OAPQD). It consists of 1,000 representative images collected from the WeChat official accounts of nine major universities in Beijing. Each image is paired with the actual page view count of the article it belonged to (ranging from 253 to 100,000 views). The set includes architecture, landscapes, people, and meeting scenes, reflecting the real-world variety of campus news.
Results and Performance
The model's performance was measured using the Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank Order Correlation Coefficient (SRCC). The study found that quality is multi-dimensional; no single feature set was sufficient to predict engagement accurately.
As shown in the results, the triple-fusion model (BRISQUE + GMLF + NFERM) achieved the best performance with a PLCC of 0.4764 and an SRCC of 0.3863. This was a massive improvement over using individual "direct use" models, which hovered near zero correlation. The results confirm a positive correlation between calculated picture quality and page views, proving that aesthetic optimization is a viable strategy for increasing digital reach.
Conclusion and Future Work
While the model is a strong step forward, the authors note that 1,000 images is a small sample size for complex deep learning architectures. Additionally, the model focuses on how an image looks (quality), but not necessarily what is in it (content). Future work aims to explore how diffusion strategies, such as the timing and platform of the post, interact with image quality to drive engagement. Publicity departments could eventually use this model as a "pre-flight" tool to test images before publication.
Summary updated