Wavelet Packet Transform and Deep Learning-based Fusion of Audio-Visual Signals: A Novel Approach for Enhancing Laser Cleaning Effect Evaluation

In response to the insufficient detection capability of laser paint stripping effects for a single modality and the high operational and cost requirements of existing multi-monitoring technologies, a method is proposed to integrate visual and auditory signals for evaluating laser paint stripping effects. Utilizing wavelet packet transformation for a more detailed understanding of the variations in paint-stripping sound signals, more representative energy features are extracted. The EfficientNetv2 network, optimized with an attention mechanism, further enhances the focus on crucial features. The image feature vectors are concatenated with the energy features extracted from the sound signals, forming a new and more informative feature vector for paint stripping effect discrimination. Experimental results demonstrate that the multi-feature fusion detection algorithm significantly improves the accuracy of paint stripping effect discrimination, reaching 98.7%. The 98.9% F1-Score and the smoothly converging loss curve also indicate the algorithm's effective control over category imbalance and training stability. This research is of paramount importance for improving the evaluation of laser cleaning technology effects and provides insights into multi-modal feature fusion for other relevant fields of study.