Self-Supervised Topologically Invariant Manifold Learning for Railway Image Quality Assessment

Tingqiong Cui; Yibu Yang; Yang Li; Jiahao Fu; Xiaoliu Luo; Xu Wang; Mengzhu Wang; Siyuan Liu

cs.CVAug 15, 2026

Self-Supervised Topologically Invariant Manifold Learning for Railway Image Quality Assessment

Authors: Tingqiong Cui, Yibu Yang, Yang Li, Jiahao Fu, Xiaoliu Luo, Xu Wang, Mengzhu Wang, Siyuan Liu,

Organizations: CRRC Chongqing Co., Ltd., Chongqing 400000, China · School of Mathematical Sciences, Chongqing University of Technology, Chongqing 400054, China · College of Computer Science, Sichuan University, Chengdu 610065, China · School of Artificial Intelligence, Hebei University of Technology, China ·

Abstract

Existing blind image quality assessment (BIQA) methods typically rely on synthetic distortions and subjective annotations, limiting generalization in real-world domains. To address this, we propose a fully self-supervised BIQA framework based on topologically invariant manifold learning under boundary constraints, which constructs a stable quality reference without manual labels. The framework generates progressive background dilution scales via repeated random cropping around each target; exploiting the monotonic degradation of target information density across these scales, it establishes a self-constrained quality manifold. A linearized spatial moment projection eliminates geometric distortions from random cropping; then a monotonicity divergence filter prunes background-sensitive evaluators, isolating an elite pool $\mathcal{M}_{\text{elite}}$ . A robust M-estimator with a principal component stabilizer fuses the metrics into an asymptotically efficient pseudo-ground truth $q_{\text{PGT}}$ , contracting variance toward the Cramér-Rao lower bound. Extensive evaluations demonstrate that the elite evaluator pool, distilled from 11 baseline metrics, secures superior zero-shot transferability across standard synthetic and wild benchmarks (CSIQ, LIVEC, LIVE-2). Concurrently, deployments on the CQU Railway Rolling Stock Surveillance Dataset (2,797 images) yield a manifold cosine similarity $>0.999$ and a 100.0% survival rate under industrial extreme stresses, robustly validating its cross-paradigm decoupling and topological resilience.

Figures & tables

Fig. 1: Overall pipeline architecture of the proposed self-supervised target-oriented BIQA framework with eigen-spectrum purification. The consensus engine coordinates its matrix streams via four interconnected analytical stages to establish a mathematically closed loop: (1) Stage 1 (Stochastic Sampling and Target Geometry Localization) localizes the rigid target geometry via a YOLO-SAM cascade and construct the invariant Sampling Field via stochastic scaling. (2) Stage 2 (Spatial Moment Linearization and Orthogonal Projection) executes an orthogonal subspace complement projection to decouple composition-driven spatial artifacts. (3) Stage 3 (Eigen-Subspace Spectrum Purification) prunes weak evaluators prone to background overfitting, isolating a coherent elite pool governed by the Perron-Frobenius conditions. (4) Stage 4 (Robust Fusion and Numerical Stability Bounds) synthesizes the final pseudo-ground truth $q_{\text{PGT}}$ via a smooth Leclerc-type M-estimator regularized by a non-parametric principal component stabilizer to sustain high structural resilience under industrial extreme singularity stresses, securing an asymptotic system survival rate of 100.0% under extreme singularity stresses, as empirically validated in Section V.

Computational Component	Deployment Specification
Central Processing Unit (CPU)	Intel Core i9-13900K
Graphics Processing Unit (GPU)	NVIDIA 1080Ti (11GB)
Random Access Memory (RAM)	36GB DDR5 Dual-Channel
Operating System	Windows 11 Professional
Core Mathematical Stack	NumPy 1.25, PyTorch 2.1

TABLE I: System Hardware and Software Computing Bounds

Algorithm Name	Agg. Energy $\uparrow$	CRLB SE $\downarrow$	Quality Gain
Our Framework	25.1419	0.0003	+9.8267
PI Baseline	24.8079	0.0032	$-25.3989$
CLIPIQA+	21.8302	0.0963	$+30.6698$
NRQM	21.0596	0.0764	$+21.3086$
NIMA	17.7060	0.0011	$-32.6198$
MUSIQ	18.2627	0.0016	$-29.2012$

TABLE II: Efficiency Analysis Across Evaluators

Projective Space Config.	Energy $\uparrow$
Our Framework	25.1419
Linear Convex Hull Bound	24.8079
Conventional Linear Combination	19.3422

TABLE III: Manifold Energy Dividend and Convex Hull Breach Comparison

Fusion Strategy Infrastructure	Agg. Gain
Our Framework (Purified)	+9.8425
Traditional Inverse-Variance	$-15.4855$

TABLE IV: Ablation of the Hard-Gate Purified Constraint Layer

Axis	Exp. Var. (%) $\uparrow$	Cum. Var. (%)
PC1	74.68	74.68
PC2	11.21	85.89
PC3	4.37	90.27
PC4	3.25	93.51
PC5	2.42	95.93
PC6	2.18	98.12

TABLE V: Scree Analysis of Eigen-Subspace Over Elite Pool

Alignment Configuration	PC1 Exp. Var. (%) $\uparrow$
Our Framework (With Lock)	74.68
Naive PCA (Without Lock)	14.25

TABLE VI: Ablation of Polarity Lock: Variance Explanation Comparison

Dataset Partition	PC1 Var.	Cosine Sim.
High-Margin Pool ( $N_{1}=924$ )	69.83%	0.998259
Constrained Pool ( $N_{2}=1,873$ )	76.36%	0.998259

TABLE VII: Topological Stability Under Heterogeneous Truncation

Stratified Spatial Tier	PC1 Var.	Cosine Sim.
Tier-1: High Abund. ( $n=700$ )	68.20%	0.999010
Tier-2: Mod. Margin ( $n=699$ )	73.77%	0.999783
Tier-3: Const. Margin ( $n=699$ )	76.27%	0.999837
Tier-4: Rel. Const. ( $n=699$ )	78.18%	0.999147

TABLE VIII: Stratified Multi-Tier Spatial Invariance Analysis

Configuration	Effective Domain	Survival $\uparrow$
Our Framework	$x\in[10^{-6},1.0]$	100.0000%
Naive Kinetics	$x\in[10^{-4},1.0]$	64.1258%

TABLE IX: Algorithmic Shielding and Numerical Stability Analysis

Hypothesis Pairwise Track	Statistic $W$	$P$ -Value
Our Framework vs. CLIPIQA+	3,884,197	$<0.0001$
Our Framework vs. NRQM	3,944,103	$<0.0001$
Our Framework vs. PI	3,964,191	$<0.0001$
Our Framework vs. NIQE	3,966,336	$<0.0001$

TABLE X: Non-Parametric Wilcoxon Signed-Rank Test Results

Validation Pairwise Track	SRCC $\uparrow$	PLCC $\uparrow$	$P$ -Value
Our P-GT vs. Independent FR-VIF	0.3184	0.2704	$<0.0001$

TABLE XI: Independent Decoupling Verification Against FR Baseline

Ablation Set Type	PLCC	SRCC
Elite Decision Pool ( $M_{e}=7$ )	1.0000	1.0000
Deep Elite Pool ( $M_{e}=3$ )	0.9842	0.9765
Clipped Elite Pool ( $M_{e}=4$ )	0.9541	0.9418

TABLE XII: Elite Ablation Test: Resiliency Under Massive Juror Loss

Algorithm Architecture	$R^{2}$ Score $\uparrow$	RMSE $\downarrow$
Our Framework (P-GT)	0.9854	1.4251
CLIPIQA+	0.9624	2.1024
NRQM	0.9415	2.6841
NIMA	0.9238	3.1092
PI	0.9105	3.5412
MUSIQ	0.8942	4.6851

TABLE XIII: Kinetics Goodness of Fit over Information Dilution Manifold

Baseline	CSIQ		LIVEC		LIVE-2
Method	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC
MUSIQ	0.6889	0.7513	0.4509	0.5259	0.5693	0.5635
NIMA	0.3516	0.3654	0.0240	0.1432	0.5326	0.5346
MANIQA	0.6345	0.6345	0.5434	0.5904	0.4935	0.5287
BRISQUE	0.5560	0.7363	0.0469	0.0641	0.3643	0.4115
CLIPIQA+	0.7245	0.7709	0.3313	0.3876	0.3349	0.3692

TABLE XIV: Cross-Dataset Zero-Shot Performance (SRCC/PLCC) Across Three Benchmarks Under Consensus Quality Evaluation

Image Quality Assessment

Image Quality Assessment measures perceived image quality, distortion, and usefulness with computational metrics and human-aligned evaluation.

See all 174 papers

SHAMISA: SHAped Modeling of Implicit Structural Associations for Self-supervised No-Reference Image Quality Assessment
Mar 14, 2026Mahdi Naseri, Zhou WangImage Quality Assessment Self-Supervised Learning
Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, ON N2L 3G1, Canada
Pixel Perfect: Relational Image Quality Assessment with Spatially-Aware Distortions
May 4, 2026Fadeel Sher Khan, Long N. Le, Abhinau K. Venkataramanan +2Image Quality Assessment
The University of Texas at Austin Austin, TX, USA · Samsung Research America Plano, TX, USA
MR-IQA: A Unified Margin View of Regression and Ranking for Blind Image Quality Assessment
Jun 29, 2026Yuan Li, Youyuan Lin, Zitang Sun +4Learning to Rank Image Quality Assessment
Graduate School of Informatics, Kyoto University

Mar 14, 2026J/K move · Enter open · S save

SHAMISA: SHAped Modeling of Implicit Structural Associations for Self-supervised No-Reference Image Quality Assessment

Mahdi Naseri, Zhou Wang

Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, ON N2L 3G1, Canada

Image Quality Assessment Self-Supervised Learning Self-Supervised Visual Representation Learning

Open paper AskPDF arXiv

No-Reference Image Quality Assessment (NR-IQA) aims to estimate perceptual quality without access to a reference image of pristine quality. Learning an NR-IQA model faces a fundamental bottleneck: its need for a large number of costly human perceptual labels. We propose SHAMISA, a non-contrastive self-supervised framework that learns from unlabeled distorted images by leveraging explicitly structured relational supervision. Unlike prior methods that impose rigid, binary similarity constraints, SHAMISA introduces implicit structural associations, defined as soft, controllable relations that are both distortion-aware and content-sensitive, inferred from synthetic metadata and intrinsic feature structure. A key innovation is our compositional distortion engine, which generates an uncountable family of degradations from continuous parameter spaces, grouped so that only one distortion factor varies at a time. This enables fine-grained control over representational similarity during training: images with shared distortion patterns are pulled together in the embedding space, while severity variations produce structured, predictable shifts. We integrate these insights via dual-source relation graphs that encode both known degradation profiles and emergent structural affinities to guide the learning process throughout training. A convolutional encoder is trained under this supervision and then frozen for inference, with quality prediction performed by a linear regressor on its features. Extensive experiments on synthetic, authentic, and cross-dataset NR-IQA benchmarks demonstrate that SHAMISA achieves strong overall performance with improved cross-dataset generalization and robustness, all without human quality annotations or contrastive losses.

Abstract

Image Quality Assessment

Explore similar work

SHAMISA: SHAped Modeling of Implicit Structural Associations for Self-supervised No-Reference Image Quality Assessment

Pixel Perfect: Relational Image Quality Assessment with Spatially-Aware Distortions

MR-IQA: A Unified Margin View of Regression and Ranking for Blind Image Quality Assessment