Flip-Robust Neural Image Assessment (FR-NIMA) foSpatially Consistent IQA
Issued Date
2026-04-01
Resource Type
eISSN
22869131
Scopus ID
2-s2.0-105037938543
Journal Title
Ecti Transactions on Computer and Information Technology
Volume
20
Issue
2
Start Page
303
End Page
318
Rights Holder(s)
SCOPUS
Bibliographic Citation
Ecti Transactions on Computer and Information Technology Vol.20 No.2 (2026) , 303-318
Suggested Citation
Tanawongsuwan R., Phongsuphap S. Flip-Robust Neural Image Assessment (FR-NIMA) foSpatially Consistent IQA. Ecti Transactions on Computer and Information Technology Vol.20 No.2 (2026) , 303-318. 318. doi:10.37936/ecti-cit.2026202.264823 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/116707
Title
Flip-Robust Neural Image Assessment (FR-NIMA) foSpatially Consistent IQA
Author(s)
Author's Affiliation
Corresponding Author(s)
Other Contributor(s)
Abstract
Neural Image Assessment (NIMA) has become a widely adopted approach for blind image quality assessment (BIQA), yet it remains sensitive to simple spatial transformations such as horizontal fiips. Such variation can lead to inconsistent predictions, even when the perceived visual content remains largely unchanged. To address this issue, we introduce Flip-Robust Neural Image Assessment (FR-NIMA), an enhanced training strategy that enhances the spatial robustness of BIQA models. Instead of modifying network architectures, FR-NIMA incorporates a fiip-consistency regularization term that penalizes discrepancies between the predicted quality distributions of an image and its horizontally fiipped counterpart. Two variants—one-branch and two-branch formulations—are explored, both introducing no additional model parameters. FR-NIMA is evaluated across four CNN backbones (MobileNetV2, VGG19, Xception, InceptionV3) and one Vision Transformer (ViT-Small) using the LIVE dataset and two additional test sets representing distinct scene types. Performance is assessed using complementary metrics, including the Test Loss (EMD<sup>2</sup>), Absolute Flip Gap (|FlipGap|), Flip-Consistency Win Rate (FCWR), Average FlipGap Delta (AFGD), and Average Flip-Gap Ratio (AFGR). Experimental results demonstrate that FR-NIMA effectively reduces fiip-gap magnitude and variability while maintaining comparable test accuracy across all backbones. These findings establish FR-NIMA as a simple yet effective framework for enhancing the stability, spatial consistency, and trustworthiness of deep IQA models.