RouteGraph-Mona: Confusion-Aware Routing Fine-Tuning for Mineral Image Classification

• Mineral image classification can support geological exploration and resource development, while conventional identification methods are costly and complicated for large-scale, low-cost automation.Image-based classification is easier to deploy, and several CNN-, feature-fusion-, DenseNet-, YOLOv8-, and Transformer-based approaches show potential.Source: §1 · pp. 1–2
• Mona’s static averaging may miss image-specific scale preferences, while standard classification-driven adaptation does not explicitly address confusion among visually similar mineral categories.Mineral categories can vary substantially within class and share local visual cues across classes.Source: §1 · p. 2
• Adaptive routing weights define a low-dimensional route space that captures image-specific scale preferences and emphasizes branch-selection patterns over appearance details.The authors argue this makes routing signatures suitable for mineral images with substantial intra-class variation.Source: §1 · p. 3
• RouteGraph-Mona replaces Mona’s static multi-scale aggregation with sample-adaptive routing and regularizes routing signatures using class-wise anchors and a confusion-weighted route margin.An online confusion graph gives larger route-margin penalties to visually confusing category pairs; the regularization is not needed during inference.Source: §1 · p. 3
• Across three public mineral datasets and two visual backbones, RouteGraph-Mona improves mean accuracy over Mona and remains competitive with representative baselines.The experiments cover three public mineral image datasets and two visual backbones.Source: §1 · p. 3

2.1 Mineral Image Classification Methods

• Mineral recognition progressed from handcrafted color, texture, and shape descriptors to deep CNN and Transformer models, often strengthened with transfer learning, fusion, augmentation, or multi-scale context.Unlike approaches centered on stronger classifiers or feature extractors, RouteGraph-Mona adapts pre-trained visual backbones and regularizes class-aware routing behavior.Source: §2.1 · p. 3

2.2 Parameter-Efficient Visual Fine-Tuning

• Parameter-efficient methods such as adapters, LoRA, and visual prompt tuning reduce fine-tuning costs, but Mona-based approaches do not explicitly model mineral category structure.RouteGraph-Mona adds class-wise route anchors and an online confusion graph to separate routes for visually ambiguous categories.Source: §2.2 · p. 4

3.1 Overall Framework

• RouteGraph-Mona replaces static branch averaging with sample-adaptive routing, then regularizes branch-selection signatures rather than visual embeddings to separate visually confusing categories.At inference, it uses a standard forward pass and predicts from classification logits.Source: §3.1 · p. 4

3.2 Parameter-efficient Fine-tuning Protocol

• The protocol freezes ImageNet-1K-pretrained ViT [7] and Swin backbones, inserts Mona modules after attention and feed-forward sublayers, and updates task-specific components while keeping the confusion graph nontrainable.Updated components include Mona projections, depthwise-convolution branches, the scale router, class-wise route anchors, and the classification head.Source: §3.2 · p. 4

3.3 Routing Signature Generation

• Each Mona module uses 3×3, 5×5, and 7×7 branches; a router fuses global context with pooled projected tokens and summarizes weights across modules into normalized signatures.The signature concatenates the mean and standard deviation of routing weights across 2L modules; uniform weights recover Mona’s static average fusion.Source: §3.3 · pp. 5–6

3.4 Class-wise Route Anchors

• Learnable class-wise route anchors act as route-space prototypes, with the main experiments using one anchor per mineral category and a route-level cross-entropy anchor loss.Sample-to-class route similarity uses the closest anchor; the formulation permits multiple anchors when finer-grained prototypes are needed.Source: §3.4 · p. 6

3.5 Online Confusion-guided RouteGraph Regularization

• An online row-normalized confusion graph is updated with momentum from detached predictions, and its confusion-weighted route margin penalizes confusing class pairs more strongly.The graph removes diagonal prediction probability and renormalizes the remaining probabilities; larger graph entries indicate stronger confusion.Source: §3.5 · p. 7

3.6 Training Objective and Inference Stage Details

• Training combines classification cross-entropy with route regularization: anchor loss guides class-level routing, while graph loss separates confusing classes; inference needs no labels, graph updates, or auxiliary losses.The adaptive router remains active at inference, and predictions use standard classification logits.Source: §3.6 · p. 7

4.1 Datasets

• Evaluation spans Minet (955 images, seven classes), MinetV2 (2,091 images, the same taxonomy), and MineralPhotos (39,354 images, 15 classes).Minet is small and imbalanced; MinetV2 has greater diversity and ambiguity, including chrysocolla–malachite and muscovite–quartz similarities. Exact-hash duplicate entries with cross-class label conflicts were removed.Source: §4.1 · p. 8

4.2 Compared Methods

• RouteGraph-Mona is compared with linear probing, full fine-tuning, adapter tuning, LoRA, Mona, and mineral classifiers DenseNet-201, EfficientNet-B0, YOLOv8-cls, and SwinMin.Mona is the direct baseline; the mineral classifiers span conventional and efficient convolutional networks, a modern classification framework, and a mineral-recognition Transformer.Source: §4.2 · p. 8

4.3 Experimental Settings

• Experiments use pretrained ViT-B/16 [7] and Swin-B, 60%/20%/20% train-validation-test splits, exact-hash grouping, and seeds 0, 666, and 2026.The splits use cleaned datasets and group exact-hash duplicates to prevent leakage.Source: §4.3 · p. 8
• Accuracy is the main metric; each run evaluates its lowest-validation-loss checkpoint once on test data, and pretrained models use ImageNet-1K weights.Fine-tuning baselines share splits, input resolution, augmentation, optimizer, early-stopping criterion, seeds, and checkpoint-selection rule; mineral baselines follow standard settings or official implementations when available.Source: §4.3 · pp. 8–9
• Unless otherwise specified, experiments use 224 × 224 inputs, AdamW, weight decay 1 × 10−4, and early-stopping patience 15.RouteGraph-Mona uses one route anchor per class, λm = 0.03, mg = 0.05, and β = 0.95.Source: §4.3 · p. 9

4.4 Main Comparison with Fine-Tuning Methods

• RouteGraph-Mona gains 1.18, 0.78, and 2.40 percentage points over Mona on Minet, MinetV2, and MineralPhotos with ViT-B/16, and 1.68, 1.42, and 2.59 points with Swin-B.These mean-accuracy gains are reported across all three datasets and both visual backbones.Source: §4.4 · p. 9
• RouteGraph-Mona outperforms linear probing, gains stably over adapter tuning and LoRA, and is competitive with full fine-tuning in most settings while keeping the pretrained backbone frozen.The comparison supports explicit route-space regularization alongside general feature adaptation for mineral image classification.Source: §4.4 · p. 9
• On cross-dataset evaluation, RouteGraph-Mona improves Mona from 73.35% to 73.98% for Minet →MinetV2 and from 94.44% to 94.61% for MinetV2 →Minet.The reported gains suggest mild robustness under dataset shift.Source: §4.4 · p. 9

4.5 Class-Balanced Evaluation and Statistical Analysis

• On MinetV2 with Swin-B, RouteGraph-Mona raises accuracy to 79.64±0.98, macro-F1 to 79.00±0.84, and mA to 78.97±1.12, gains of 1.42, 1.64, and 1.59 percentage points over Mona.These metrics use the same three seeds as Table 1.Source: §4.5 · p. 9
• Matched-seed comparisons show positive gains, but the accuracy paired t-test is not significant (p = 0.151), so the results support a trend rather than statistical significance.The comparison uses a small number of seeds.Source: §4.5 · p. 10

4.6 Accuracy–Efficiency Trade-off

• Under ViT-B/16 and Swin-B, RouteGraph-Mona improves accuracy over Mona while retaining a low trainable-parameter ratio, but incurs additional training-time memory.The performance gain does not rely on large-scale backbone unfreezing; Figure 3 compares accuracy, trainable-parameter ratio, and peak GPU memory.Source: §4.6 · p. 10, Figure 3 · p. 10
• RouteGraph-Mona updates lightweight adaptation components, scale routers, route anchors, and the classification head without large-scale backbone updates.At inference, the scale router remains active, but no labels, graph updates, or auxiliary losses are needed; predictions use classification logits.Source: §4.6 · p. 10

4.7 Comparison with Mineral Image Classification Methods

• RouteGraph-Mona achieves 93.94%, 79.64%, and 81.09% accuracy on Minet, MinetV2, and MineralPhotos, beating SwinMin by 2.53, 0.63, and 0.77 percentage points, respectively.It achieves the best accuracy among the compared mineral-classification baselines on all three datasets.Source: Table 2 · p. 11, §4.7 · p. 11
• The mineral-classification comparison includes DenseNet-201, EfficientNet-B0, YOLOv8-cls, and SwinMin, spanning convolutional, modern classification, and mineral-specific Transformer baselines.The comparison covers three mineral datasets.Source: §4.7 · p. 11, Table 2 · p. 11

4.8 Ablation Study

• Adaptive routing brings small, consistent gains on Minet and MinetV2, while route anchors slightly improve MinetV2 accuracy and reduce Minet’s standard deviation.The findings suggest that sample-adaptive routing supports route-space modeling and category-level route prototypes regularize routing signatures.Source: §4.8 · pp. 11–12
• 78.38% to 78.77% on MinetV2 is the accuracy gain from adding the confusion-guided route margin, with stronger gains where visual ambiguity is greater.Source: §4.8 · p. 12
• 93.94% on Minet and 79.64% on MinetV2 are RouteGraph-Mona’s highest accuracies among the ablation variants.The study attributes the combined gains to adaptive routing defining the route space, anchors providing category-level references, and the margin emphasizing separation between visually confusing categories.Source: §4.8 · p. 12

4.9 Feature-space versus Route-space Regularization

• Center Loss slightly decreases accuracy on both datasets, while supervised contrastive learning improves MinetV2 but decreases Minet.These results show that generic feature-level discriminative losses do not consistently benefit mineral image classification.Source: §4.9 · p. 12
• Feature Anchor+Graph applies RouteGraph-Mona’s anchor, confusion-weighted margin, and online graph constraints to visual feature embeddings instead of routing signatures.This design isolates the effect of the regularization space.Source: §4.9 · p. 12
• Feature Anchor+Graph performs worse than RouteGraph-Mona on both datasets, supporting the benefit of applying structural constraints to adaptive multi-scale routing signatures rather than visual embeddings.The matched variant uses the same anchor-based prototype constraint, confusion-weighted margin, and online graph construction.Source: §4.9 · p. 12

4.10 Effect of Confusion Graph Construction

• 79.64±0.98% accuracy is achieved by the online confusion graph, improving the no-graph setting by 1.18 percentage points with the lowest variance among evaluated graph strategies.Prediction-derived confusion weights route-margin penalties toward category pairs more likely to be visually confused.Source: Table 5 · p. 13, §4.10 · p. 13
• 78.46±1.71% accuracy for no graph, 78.30±2.86% for uniform weighting, and 78.69% with 2.61 standard deviation for random weighting show no stable benefit from fixed or random graphs.Random weighting also lacks semantic correspondence to actual category confusion.Source: Table 5 · p. 13, §4.10 · p. 13

4.11 Hyperparameter Sensitivity

• One route anchor per class already provides strong mean accuracy, while increasing K yields no consistent improvement and performance notably drops at K = 5.The sensitivity analysis uses MinetV2 with the Swin-B backbone; overlapping error bars mean these are overall trends rather than statistically significant differences.Source: §4.11 · pp. 13–14, Figure 4 · p. 13
• RouteGraph-Mona decreases total off-diagonal confusion error by 15.3% relative to Mona and widens the separation gap between intra-class and confusing inter-class route distances.Negative confusion-change values indicate reduced off-diagonal errors.Source: Figure 5 · p. 14
• The default λm = 0.03 achieves competitive performance, while other evaluated regularization weights show only moderate fluctuations.Increasing regularization strength does not consistently yield additional gains.Source: Figure 4 · p. 13, §4.11 · p. 14
• The route margin mg = 0.05 performs well, whereas increasing it to 0.07 leads to a performance decrease.The evaluated settings show overall sensitivity trends, not statistically significant differences, given overlap among several error bars.Source: Figure 4 · p. 13, §4.11 · p. 14

4.12 Analysis of Route-Space Geometry

• RouteGraph-Mona reduces total off-diagonal confusion error by 15.3% versus Mona and increases confusing inter-class route distance from 0.862 × 10−2 to 1.206 × 10−2.Figure 5 analyzes class confusion and route-space geometry.Source: §4.12 · p. 14
• The route-space separation gap widens from 0.076 × 10−2 to 0.144 × 10−2, becoming 1.90× Mona’s gap.This suggests improvement mainly reflects better separation of visually confusing categories rather than simply compressing samples from the same class.Source: §4.12 · p. 14

• Across three mineral image datasets and two visual backbones, RouteGraph-Mona improves mean accuracy over Mona and performs competitively with representative fine-tuning and mineral image classification baselines.Source: §5 · p. 15
• RouteGraph-Mona replaces Mona’s static multi-scale aggregation with sample-adaptive routing and constrains routing signatures using class-wise route anchors and an online confusion-guided route margin.It regularizes routing signatures rather than visual feature embeddings directly, improving relative route-space separation between visually confusing mineral categories.Source: §5 · p. 15
• Experiments find one route anchor per class empirically sufficient, while future work may explore class-adaptive anchor allocation and more robust confusion-graph updates on larger, more diverse mineral datasets.Source: §5 · p. 15