The widespread application of the drill-and-blast method in tunnel construction generates instantaneous high-concentration dust, posing severe threats to workers’ health and safety. However, existing contact-based monitoring techniques, such as filter membrane weighing and light-scattering sensors, are hindered by operational complexity, maintenance challenges, data latency, and an inability to capture whole-field distribution. To address these limitations, this study conducted in situ experiments to construct a dust image dataset containing rich “real-world noise”. Analysis revealed significant variations in visibility, contrast, and light-scattering intensity across different concentration levels, establishing a physical basis for visual feature extraction. Consequently, a dust concentration prediction method based on transfer learning and CNN was proposed. Comparative training of six network models, including ResNet50, demonstrated that VGG16 achieved the best performance with an R² of 0.9684, verifying the method’s feasibility. Furthermore, hyperparameter optimization (learning rate: 0.001; batch size: 32; dataset split: 8:1:1) was shown to further enhance prediction accuracy.