Intelligent Cover Image Selection System for News Articles: Image Quality Assessment and Smart Cropping

This article presents an intelligent cover image selection system developed for NetEase News app, addressing the challenge of automatically selecting optimal cover images from article illustrations. The system consists of two main components: candidate image generation and image cropping.

For candidate image generation, the system first filters unsuitable images including QR codes, images with excessive whitespace, images with too much text, images with competitor logos, and similar images. It employs deep convolutional neural networks for QR code detection, pixel statistics for whitespace analysis, text detection and recognition algorithms for text-heavy images, and grayscale histogram comparison for duplicate detection.

The image quality assessment model uses a novel pairwise comparison training approach rather than traditional point-wise training. The model considers multiple dimensions including image clarity, color saturation, contrast, brightness, subject emphasis, and composition complexity. By inputting image pairs and penalizing ranking inversions, the model better aligns with human subjective evaluation.

For image cropping, the system employs an object detection-based approach that predicts croppable regions for different aspect ratios. The model extends classic object detection algorithms (Faster RCNN, FRCNN, SSD, YOLO) by adding multiple output channels, each sensitive to a specific aspect ratio range. This allows the system to automatically detect the optimal crop region while maintaining visual aesthetics.

The system has been deployed in production, significantly reducing manual editing workload and accelerating the review process. With confidence scoring, high-confidence selections can bypass manual review entirely.