Image Quality Assessment Techniques and Their Application in 58.com Recruitment Image Filtering

Image quality assessment (IQA) is a fundamental technology for evaluating the visual quality of images, which impacts information expression and user experience. IQA methods are categorized into full‑reference (FR), reduced‑reference (RR), and no‑reference (NR) approaches, each requiring different levels of reference information.

Typical IQA datasets include LIVE, TID2013, KonIQ‑10k, CSIQ, and others, providing reference images, distorted versions, and subjective scores (MOS/DMOS). Evaluation metrics such as PLCC, SRCC, PSNR, and SSIM are used to measure the correlation between algorithm predictions and human judgments.

Traditional FR methods like MSE, PSNR, and SSIM compute pixel‑wise differences, while NR methods such as BRISQUE, MSDD, NR‑DBCNN, hyperIQA, and RankIQA rely on statistical features or deep learning.

Recent CNN‑based IQA models have achieved significant performance gains. WaDIQaM uses a VGG‑16 twin network to extract features from reference and distorted patches, fusing them via concatenation and applying L1 loss. DBCNN employs twin VGG‑16 branches, a synthetic distortion branch (S‑CNN) pretrained on large datasets, and bilinear pooling for feature fusion. hyperIQA introduces a hyper‑network that adapts model weights per image, combining multi‑scale ResNet‑50 features, local distortion awareness (LDA), and a perception‑rule learning network.

The business background is 58.com’s recruitment platform, where employers upload images to a gallery. Poor‑quality images (blurred, noisy, low resolution, excessive text, logos, certificates, etc.) degrade user experience, with an initial bad‑image rate of 9%.

To address this, a customized IQA pipeline was built: an IQA model provides a base quality score, which is then corrected using OCR, background clustering, resolution checks, aspect‑ratio checks, and text‑ratio analysis. Scores are clustered with K‑means and thresholds are tuned to map continuous predictions to three quality categories (low, medium, high).

Models (WaDIQaM, DBCNN, hyperIQA) were trained on synthetic datasets (LIVE, TID2013) and the authentic Koniq‑10k dataset. hyperIQA trained on Koniq‑10k achieved the highest accuracy (86.93%) on the 58zhaopin‑5k test set. Replacing ResNet‑50 with ResNet‑18 reduced model complexity while maintaining performance.

After applying the score‑correction rules, overall accuracy rose to 94.72%, and the deployed service processes ~120,000 new images daily and 200,000 stored images, reducing the bad‑image rate to 0%.

The solution demonstrates how academic IQA research can be adapted to real‑world business needs, and future work will focus on making the algorithm more generalizable, capable of detecting both high‑level semantic and low‑level quality issues across various scenarios such as video quality monitoring.

References

[1] https://blog.csdn.net/Image_test/article/details/52036873?locationNum=2&fps=1 [2] V. Hosu et al., "Koniq‑10k: An ecologically valid database for deep learning of blind image quality assessment." [3] S. Su et al., "Blindly Assess Image Quality in the Wild Guided by A Self‑Adaptive Hyper Network." [4] S. Bosse et al., "Deep neural networks for no‑reference and full‑reference image quality assessment." [5] W. Zhang et al., "Blind image quality assessment using a deep bilinear convolutional neural network." [6] https://blog.csdn.net/caoleiwe/article/details/49045633. [7] https://www.cnblogs.com/zhangzizi/p/14734071.html