Objective Laser paint removal is widely applied because it is non-contact, locally controllable, and environmentally friendly. During laser paint removal, insufficient cleaning can leave coating residues that affect subsequent repainting, whereas excessive cleaning may cause substrate erosion or oxidation. Therefore, monitoring paint-removal cleanliness and substrate integrity is a key measure for achieving residue-free and damage-free paint removal. Existing laser-cleaning monitoring methods often struggle to achieve both large-area coverage and high sensitivity. To overcome this limitation, based on the speckle-imaging principle of coherent scattering and by exploiting the speckle-image differences induced by subtle surface-structure changes, a speckle-based monitoring method with high sensitivity to surface-structure variations is proposed.

Methods The experimental workflow includes speckle image acquisition, speckle image filtering, feature extraction, feature selection and classification-model training. First, speckle images of specimens under different paint-removal states are acquired with a camera, labeled, and organized into a dataset. A three-class dataset (Clean/Residue/Damage) containing 630 speckle images (210 per class) was built. For each class, 150 images were used for training and cross-validation, while two independent test sets were prepared for pronounced and slight states (30 images per class in each set). Bilateral filtering is applied to the raw speckle images for noise suppression, thereby improving feature stability (Fig.2). Next, features are extracted from multiple perspectives, including statistical, texture, and scale-related descriptions, to fully characterize the information contained in speckle images. The Spearman correlation coefficient is then used for feature selection (Fig.3). Furthermore, Linear Discriminant Analysis (LDA) is employed to project the selected features into a discriminant space to verify separability among different surface states (Fig.4). Finally, the selected features are fed into multiple machine-learning classifiers (including SVM, KNN and RF) for training and comparison, and cross-validation together with evaluation metrics are used to determine the optimal model as the final recognition model (Tab.1, Tab.2).

Results and Discussions The features retained after Spearman-based selection exhibit clear separability in the LDA discriminant space (Fig.4), indicating that the selected feature subset can effectively represent differences among paint-removal states as reflected by speckle patterns. Comparative results among classification models show that SVM achieves the best overall performance and is therefore selected as the final classification model (Tab.1, Tab.2). For pronounced coating-residue and pronounced substrate-damage states that can be identified by a vision-based scheme, the model achieves an identification accuracy of 100%, with no false positives or missed detections (Fig.7, Tab.3). For slight coating-residue and slight substrate-damage states that are difficult to identify by the vision-based scheme, the model still achieves an overall classification accuracy of 91% (Tab.4). The macro-averaged AUC is 0.992, and the per-class recalls are 100% (slight residue), 77% (clean), and 97% (slight damage) (Fig.10). Furthermore, when the speckle size was expanded to 25 mm for testing, the model achieved an Accuracy, Macro-Precision, Macro-Recall, and Macro-F1 of 1.00 on the independent test set for slight states, indicating that under large-area conditions, the method maintains high-sensitivity recognition capability for slight states. Most misclassifications occur between the Clean class and slight substrate damage, as these boundary states exhibit more similar speckle statistics due to subtle and mixed surface microstructures. These results demonstrate that the proposed method is applicable not only to pronounced-state recognition but also enables effective discrimination of slight boundary states. Overall, the results confirm that the proposed speckle-based monitoring method can achieve high-sensitivity recognition under large-area conditions.

Conclusions To address the challenge that current laser cleaning monitoring methods cannot simultaneously provide large-area coverage and high sensitivity, a laser paint removal monitoring method based on speckle imaging and machine learning is developed. The method enhances feature stability through bilateral filtering, represents speckle images using multi-level features, performs redundancy removal and feature selection via Spearman correlation analysis, verifies feature separability using LDA, and finally realizes classification of paint-removal states using an SVM model. Experimental results show that the recognition accuracy is 100% for vision-identifiable pronounced coating-residue and pronounced substrate-damage states, and remains 91% for vision-difficult slight coating-residue and slight substrate-damage states. The further test with speckle size expanded to 25 mm yielded perfect metrics (1.00) for slight states, confirming that the proposed method combines large-area coverage with high-sensitivity recognition capability. This method can be extended to cleaning scenarios involving different substrates and composite coatings, and provides an effective technical means for closed-loop control of laser cleaning.