Selective Intensity Gating Ensembles Improve Architectural Image Classification Under Corruption

A newly developed computer vision architecture termed the Multi-Threshold Selective Ensemble achieves marked gains in classification robustness under image corruptions and real distribution shifts when identifying architectural landmarks. In experimental evaluations, the model produced statistically significant performance improvements of 2.4 to 6.3 percentage points over baseline approaches during corrupted test conditions, accompanied by odds ratios ranging between 3.7 and 4.2 and Holm-adjusted p-values below 0.004. At the same time, the researchers openly established that while their technique attains high accuracy on clean benchmark imagery, the margin of improvement over a competitive single-model baseline under standard uncorrupted conditions remains modest and statistically insignificant. Consequently, the primary validated merit of the framework rests in its resilience to severe visual degradation rather than nominal clean-set accuracy gains.

The investigation centers on the fine-grained classification of architectural monuments, an operational domain recognized as an exceptionally demanding computer vision challenge. Buildings belonging to a common architectural tradition frequently display only minimal distinctions in subtle decorative motifs and fine-grained surface textures, while their captured visual appearance shifts drastically across wide illumination variations and lighting intensity ranges. Conventional classification workflows frequently attempt to mitigate these visual discrepancies by applying a single, invariant preprocessing step across an entire image repository or by constructing multi-model ensembles founded on the unverified presumption that every incorporated member provides positive analytical value. According to the study, such prevailing strategies fail to accommodate the continuous spectrum of illumination intensities and simultaneously introduce substantial computational inefficiency through the inclusion of unhelpful, redundant constituent models.

To overcome these dual limitations, the authors formulated the Multi-Threshold Selective Ensemble around two complementary algorithmic stages. The initial component relies on a nonlinear multi-threshold intensity-gating operator designed to systematically attenuate above-threshold pixel intensities scaled within the standard unit interval. By selectively dampening high-intensity regions across multiple threshold settings, the operator transforms individual source photographs into distinct, complementary image variants that are each specialized for specific segments of the intensity continuum. In their technical analysis, the authors established the formal mathematical behavior of this gating mechanism, deriving an exact identity for local image contrast and rigorously proving that the transformation preserves the bounded numeric range of the original input data throughout processing.

The second component of the architecture translates these specialized image variants into an efficient ensemble classifier. The researchers train an individual YOLO11n-cls neural network model exclusively on each variant generated by the intensity-gating operator. Rather than deploying every trained classifier simultaneously in a fixed combination, the framework executes an exhaustive subset selection procedure across validation data. This search systematically determines and retains only the exact collection of models that collectively maximizes accuracy on a held-out validation set. Any candidate model that fails to provide an empirical performance contribution to the aggregate output is pruned from the deployed system, eliminating unnecessary computational overhead while safeguarding overall classification fidelity.

Empirical validation of the system was conducted across two distinct benchmark image collections, with statistical significance rigorously verified using McNemar’s test. When evaluated on the Samarkand v2 dataset, which comprises 1,383 photographs representing twelve Central Asian architectural landmarks, the Multi-Threshold Selective Ensemble achieved an overall accuracy of 99.24 percent, bounded by a 95 percent confidence interval extending from 97.86 to 99.89 percent, alongside an F1 score of 99.25 percent. When tested on the Oxford 5k Buildings benchmark, the architecture recorded an accuracy rate of 96.58 percent. Despite these exceptionally high nominal values, comparative testing revealed that the performance differential between the full ensemble and a single optimized baseline model on clean inputs remained small and lacked statistical significance.

The true operational strength of the selective ensemble became apparent when the test imagery was exposed to synthetic corruptions and real-world distribution shifts. Under these challenging environmental conditions, the classification advantage provided by the gated ensemble widened substantially, demonstrating improvements between 2.4 and 6.3 percentage points that satisfied formal statistical thresholds with Holm-adjusted p-values below 0.004. Furthermore, the systematic pruning process identified an identical optimal model configuration across both benchmark collections, specifically settling on the three-member subset composed of MO, ME100, and ME150. The highest threshold candidate, designated as ME225, failed to provide distinct predictive utility and was consistently eliminated during validation pruning.

In addition to sustained classification performance under adverse image conditions, the selective pruning strategy delivered measurable practical advantages in computational resource expenditure. By discarding non-contributing neural models and retaining only the optimal three-member team, the architecture reduced overall inference costs by 25 percent in comparison to an unpruned fixed ensemble. The authors emphasized that the operational scope of their research is explicitly restricted to architectural imagery, positioning the investigation as a reproducible and statistically grounded examination of the specific circumstances in which multi-threshold gated ensembles provide demonstrable utility. Through this structured evaluation, the paper details precisely how tailored intensity preprocessing and disciplined ensemble minimization jointly bolster vision systems against visual corruption.

Subject of Research: Architectural image classification using multi-threshold intensity gating and selective ensemble learning

Article Title: Multi-threshold intensity gating with selective ensemble learning for robust architectural image classification

Article References: Ismatovich, A. A., Tashmuratovich, E. U., Abdualimovich, H. U., & Turdibayevich, K. R. (2026). Multi-threshold intensity gating with selective ensemble learning for robust architectural image classification. Multimedia Tools and Applications, 85(11), Article 843. https://doi.org/10.1007/s11042-026-22001-2

Image Credits: AI Generated

DOI: 10.1007/s11042-026-22001-2

Keywords: architectural image classification, multi-threshold intensity gating, selective ensemble learning, computer vision, YOLO11n-cls, image robustness, McNemar test, distribution shift, Multi-threshold, intensity, gating, selective