Type-Balanced Federated Learning for Visual Analog Meter Reading

• Sites may have limited labeled data and visual diversity, while ownership or privacy requirements can prevent pooling raw inspection images.The passages note that sites may contribute complementary meter types and imaging conditions.Source: §1 · p. 1
• FedAvg can give data-rich instrument types most of the weight, while uniform client weights do not necessarily balance types; a common federated benchmark is also lacking.The paper calls for reproducible client partitions and evaluation of both segmentation and downstream readings.Source: §1 · p. 1
• Type-balanced aggregation improves average pointer IoU over FedAvg and local training across three initializations, and reduces cross-client IoU variation relative to FedAvg.With synthetic initialization, it achieves the highest average pointer IoU (0.576) and lowest client STD (0.153) among the compared federated methods.Source: §1 · p. 1

• Conventional meter readers combine localization, pointer detection, scale extraction, and geometric interpolation, while learning-based readers use detection, segmentation, or keypoint networks.Synthetic-data training and VLM-based reading are discussed as related approaches; the paper retains explicit tick and pointer geometry for comparison with VLM readers.Source: §2 · p. 1
• Prior federated visual learning includes medical imaging and digital-register recognition, whereas this task requires dense tick/pointer segmentation and geometric decoding of continuous analog readings.The analog-meter setting also involves heterogeneous instrument types across clients.Source: §2 · p. 1
• FedAvg uses sample-size weights, FedProx constrains local updates, FedBN retains client-specific normalization, FedSAM uses sharpness-aware optimization, and q-FFL emphasizes high-loss clients.The paper's rule instead assigns each declared instrument type a fixed aggregate share, independent of its data volume or number of clients.Source: §2 · p. 2

3.1. Meter-Reading Pipeline

• A four-stage pipeline localizes the dial, segments ticks and pointer, unwraps the mask into a polar strip, and decodes the reading by tick counting and interpolation.The detector is trained separately and fixed; polar unwrapping and decoding have no trainable parameters.Source: §3.1 · p. 2
• Polar unwrapping uses the tick-pixel centroid as dial center and transforms the predicted mask into a linear strip, starting at the largest gap between neighboring ticks.For decoding, tick and pointer pixel counts form angular signals; grouped above-threshold columns provide their positions for interpolation.Source: §3.1 · p. 2

3.2. Type-Balanced Federated Aggregation

• Clients train the segmentation model locally and share model parameters; images and annotations remain local.Local training uses class-weighted cross-entropy, photometric augmentation, and light Gaussian input noise.Source: §3.2 · p. 2
• The server gives each instrument type an equal total aggregation weight and weights clients within a type by training-set size.Each type therefore contributes 1/|T| of the aggregate, regardless of its number of clients or images; the section's figure describes metadata-driven pseudo-client construction.Source: §3.2 · p. 3, Figure 2 · p. 3

4.1. The MeterFL Benchmark

• MeterFL contains 1,382 mask-annotated images organized into pseudo-clients formed from instrument-by-domain cells, with near-duplicate groups kept within one split.The benchmark’s pseudo-clients are derived from visible attributes and deterministic rules; dHash controls train/test near-duplicate leakage.Source: §4.1 · p. 3

4.2. Experimental Setup

• Evaluation uses eight clients, 1,102 training images, and 272 test images for segmentation, plus 269 GT-decodable images for complete-pipeline reading.Models train for 60 rounds of five local epochs over three seeds; the setup reports average pointer IoU, client IoU STD, median %FS, within-5%FS share, and Acc.Source: §4.2 · p. 3

4.3. Main Results: Centralized versus Federated Training

• 1.4%FS median and 87.2% Acc are achieved by the federated framework, compared with 2.1%FS for pooled LoRA-SFT and 2.4%FS and 77.0% Acc for the PRI-style variant.The framework’s only learned component is a 30.6M-parameter U-Net, trained without pooling images.Source: §4.3 · p. 4, Table 1 · p. 4
• 0.576 IoU and 0.153 STD are the best federated results in both columns for type balancing, which reduces the four pressure-gauge clients’ aggregate share from 96% to 25%.The comparison covers eight clients and 272 test images; type balancing adds no communication.Source: §4.3 · p. 4, Table 1 · p. 4
• 87.2% TB Acc is not significantly different from centralized training’s 87.9%, while the share within 5%FS is significantly higher for TB than FedAvg, 84.1% versus 80.4%.The paired 95% confidence interval for Acc includes zero; for the within-5%FS difference it is [+0.3, +7.3].Source: §4.3 · p. 4, Table 1 · p. 4

4.4. Ablations and Sensitivity

• Across three initializations, TB has the highest pointer IoU, 0.029–0.103 above the next method; end-to-end Acc varies less, from 83.6–87.2% across methods and initializations.Controls also report that type-balanced sampling improves pooled training, TB reaches similar average IoU without pooling, and retaining the dial crop and using U-Net improve K=20 pointer IoU over their alternatives.Source: §4.4 · p. 4, Figure 4 · p. 4, Figure 5 · p. 4

• Type-balanced federated learning improves pointer segmentation over FedAvg and local training while reducing cross-client performance disparity.Source: §5 · p. 4
• The framework achieves performance close to centralized training without sharing client images.Source: §5 · p. 4
• The work presents a federated framework for visual analog meter reading.Source: §5 · p. 4