ProgressIQA: Progressive Curriculum and Ensemble Self-Training for Filter-Altered Image Quality Assessment
2026 IEEE International Conference on Acoustics, Speech, and Signal Processing (IEEE ICASSP 2026)
4 – 8 May, 2026
Barcelona, Spain
[PDF]
MohammadAli Hamidi, Hadi Amirpour, Christian Timmerer, Luigi Atzori
Abstract: Filter-altered images are increasingly prevalent in online visual communication, particularly on social media platforms. Assessing the relevant perceived quality is essential for effectively managing visual communication. However, the perceived quality is content-dependent and non-monotonic, posing challenges for distortion-centric Image Quality Assessment (IQA) models. The Image Manipulation Quality Assessment (IMQA) benchmark addressed this gap with a dual-stream baseline that fuses filter-aware and quality-aware encoders via an MS-CAM attention module. However, only eight of the ten dataset folds are publicly released, making the task more data-constrained than the original 10-fold protocol. To overcome this limitation, we propose ProgressIQA, a data-efficient framework that integrates ensemble self-training, label distribution stratification, and multi-stage progressive curriculum learning. Fold-specific models are ensembled to generate stable teacher predictions, which are used as pseudo-labels for external filter-augmented images. These pseudo-labels are then balanced through stratified sampling and combined with the original data in a progressive curriculum that transfers knowledge from coarse to fine resolution across stages. Under the restricted 8-fold protocol, ProgressIQA achieves PLCC 0.7082 / SROCC 0.7107, outperforming the IMQA baseline (0.5616 / 0.5486) and even surpassing the original 10-fold evaluation in SROCC (0.7253 / 0.6870).
