A Lightweight SE-Enhanced Deep Regression Model for Reliable Analog Gauge Reading in Industrial Environments

Accurate and automated analog gauge reading is critical in industrial environments, where manual interpretation can lead to safety hazards, operational delays, and financial losses. Existing CNN- and SE-based models often struggle under real-world conditions such as illumination variation, noise contamination, and diverse gauge layouts, particularly for complex gauges (Types B and C). To address these challenges, this paper presents RegSENet, a regression-focused deep learning architecture that integrates Convolutional Neural Networks (CNNs) with Squeeze-and-Excitation (SE) blocks for enhanced channel-wise feature recalibration. By dynamically emphasizing discriminative visual features, RegSENet enables robust prediction across original, contrast-enhanced, and noise-injected images. Compared to a baseline CNN, RegSENet reduces Gauge Type A error from MSE = 0.0021 to 0.0009 (≈57%) and from MAE = 0.0456 to 0.0243 (≈46.7%). It also improves performance on visually more challenging gauges, reducing MSE from 0.0314 to 0.0208 on Type B and from 0.2981 to 0.1947 on Type C, while maintaining stable behavior under visual degradations. Indicative latency measurements on desktop-class hardware show only a small overhead relative to the baseline CNN. Geometric rules are employed exclusively for offline ground-truth generation, whereas inference is performed end-to-end via deep regression without explicit geometric computation. In addition to standard train–test evaluation, five-fold grouped cross-validation and leave-one-gauge-type-out experiments are conducted to assess generalization under domain shift, indicating consistent performance trends despite the limited dataset size. Overall, the results highlight RegSENet’s improved robustness, controlled generalization behavior, and edge feasibility in principle for industrial monitoring scenarios, subject to future embedded-device validation.