Sandblasting robots play a key role in industrial surface treatment. The accuracy of their visual recognition systems in detecting minute defects directly impacts processing quality. Improving detection accuracy requires coordinated improvements in optical system design, image processing algorithms, multimodal data fusion, and environmental adaptability optimization. To address interference from strong reflections and abrasive splash generated during sandblasting robot operations, the imaging system's hardware configuration must be optimized. A high dynamic range (HDR) industrial camera combined with a coaxial light source can be used. By adjusting the light source's angle and intensity, tiny defects can be clearly contrasted against a complex background. For example, when using a ring-shaped LED light source, controlling the incident angle between 30° and 45° can effectively suppress interference from diffuse reflections from the sandblasted surface.
The image preprocessing stage requires targeted solutions to the noise issues introduced by the sandblasting process. Traditional filtering algorithms tend to blur defect edges, while an improved algorithm based on non-local means (NLM) analyzes image block similarity to reduce noise while preserving the detailed features of tiny cracks. Combining morphological opening and closing operations can eliminate isolated noise caused by abrasive particle attachment and repair local image breaks caused by sandblasting impact. To address the blurring issues associated with dynamic sandblasting robot operations, motion compensation technology based on optical flow is introduced. By analyzing pixel displacement between consecutive frames, it reverses image blur and improves edge clarity of 0.1mm scratches by over 40%.
The application of deep learning algorithms offers new insights for micro-defect detection. Convolutional neural networks (CNNs) automatically learn the complex patterns of sandblasted surface defects through multi-layer feature extraction. To address the challenge of limited sample training, a transfer learning strategy is employed. Model parameters pre-trained in similar industrial scenarios are loaded, and only the last few fully connected layers of the network are fine-tuned, enabling the model to quickly adapt to the inspection requirements of sandblasted workpieces. Improved networks incorporating attention mechanisms, such as the Convolutional Block Attention Module (CBAM), can dynamically focus on defect areas and reduce the interference of background noise on classification results. Experimental results show that this method improves the detection rate of 0.05mm pits by 25% compared to traditional algorithms.
Multimodal data fusion is key to improving inspection robustness. By combining the sandblasting robot's visual information with laser contour scanning data, the laser point cloud is used to obtain the workpiece's 3D topography. A point cloud registration algorithm is then used to eliminate errors in visual inspection caused by variations in sandblasting thickness. For example, when the vision system detects a surface anomaly, the laser scanning module is triggered to perform precise measurements of the suspected area. By comparing the deviation between the actual contour and the theoretical model, the defect's authenticity is confirmed, effectively distinguishing process traces from actual defects and reducing false alarms. Furthermore, integrating contact data from force control sensors can further validate visual inspection results, forming a hybrid "visual-force" inspection system.
Environmental adaptability optimization requires consideration of factors such as vibration, temperature, and humidity fluctuations in the sandblasting workshop. During equipment installation, an active vibration isolation platform is configured for the sandblasting robot's vision system. Air bearing support and feedback control suppress vibrations above 6 Hz, ensuring imaging stability. The lighting system utilizes stroboscopic control technology to synchronize the light source frequency with the camera frame rate, preventing artifacts caused by workpiece movement. A temperature and humidity compensation module monitors environmental parameters in real time and automatically adjusts light source brightness to maintain illumination uniformity within ±3%, minimizing detection errors caused by environmental fluctuations.
The detection algorithm requires a dynamic threshold model to automatically adjust defect detection criteria based on sandblasting process parameters. For example, as the abrasive grit increases, the surface roughness tolerance is appropriately relaxed to prevent misidentification of normal process traces as defects. As the sandblasting pressure increases, the crack depth detection threshold is raised to avoid false positives caused by surface deformation due to impact. By establishing a mapping library between process parameters and detection thresholds, the system possesses adaptive adjustment capabilities, improving detection reliability under complex working conditions.
System validation and continuous optimization are long-term measures to ensure detection accuracy. A golden sample library containing various typical defects is established. The sandblasting robot vision system is calibrated and tested weekly to record and analyze false detections and missed detections. The algorithm model is then updated through iterative training. A visual operation interface is developed to allow operators to fine-tune inspection parameters such as sensitivity and defect type filtering according to actual needs. At the same time, all inspection data is recorded for quality traceability, forming a closed-loop management system of "inspection-feedback-improvement" to ensure the long-term stable operation of the system.
