Atomic force microscopy faces enduring challenges in imaging speed, accuracy, and automated analysis as microstructure characterization becomes increasingly critical across materials science and biomedical research. Traditional scanning methods struggle with environmental interference, causing height deviation and structure distortion, while manual analysis workflows limit throughput and introduce measurement variability. Published research in Procedia Computer Science introduces comprehensive intelligent processing systems integrating deep learning architectures with optimized scanning strategies to transform AFM data acquisition and analysis capabilities.
The analytical foundation addresses fundamental scanning limitations through trajectory optimization that enhances imaging speed while reducing deformation. Research details smooth sinusoidal curve implementation, replacing standard triangular wave trajectories to eliminate velocity discontinuities that trigger piezoelectric scanner resonance. Fourier transform analysis validates effectiveness, demonstrating optimized trajectories reduce 70Hz and 90Hz component amplitudes from 0.1655µm and 0.1001µm to 0.0032µm and 0.0014µm, respectively, at 10Hz scanning frequency, significantly suppressing high-frequency resonance and expanding scanning bandwidth for improved stability.
Complementing trajectory optimization, robust correction mechanisms address image distortion from vertical drift and false slope through improved line-fitting methodologies. The framework implements two-stage processing, combining error signal analysis for data screening with sparse sampling consistency algorithms for accurate fitting. By analyzing laser error signal correlation with topography data and applying SPASAC methods to eliminate outlier interference, the system effectively compensates image distortion and restores true surface morphology, reducing tilt artifacts from 6° to near-zero while maintaining measurement reliability across complex sample geometries.
Target region extraction leverages deep learning models integrating multi-level information through improved U-shaped network architectures. Research incorporates cross-scale information interaction strategies, enhancing feature fusion between network layers, combined with channel and spatial attention mechanisms, strengthening important information extraction. Global information guidance through dilated spatial pyramid pooling ensures high-level semantic preservation during upsampling. Experimental validation on constructed AFM datasets demonstrates superior performance, achieving 0.976 Dice coefficient, 0.954 IoU, and 0.028 MAE compared to traditional Otsu algorithms and standard deep learning approaches, processing images at approximately 20 frames per second.
Practical implementation validates methodologies through morphological analysis applications spanning polymer samples, bacterial specimens, and E. coli cell imaging. Automated measurement experiments extract comprehensive morphological features, including actual size, boundary length, footprint, shape proportion, and structure compactness, demonstrating superior accuracy compared to manual annotation benchmarks while significantly reducing analysis time and improving consistency across diverse sample types.
This research originates from Minkang Zhang, holding a Master of Science in Computer Science from the University of Southern California and a Bachelor of Science in Electrical and Computer Engineering, Cum Laude, from The Ohio State University. Professional specialization encompasses medical device software development with sustained contributions to the IH-500 immunohematology testing system, machine learning implementation utilizing CNTK and OpenCV for automated recognition, and full-stack engineering leveraging DevOps practices. Technical proficiency spans Python, Java, C#, and deep learning frameworks, with project achievements including 95% accuracy neural network development and advanced AI agent implementation.
These contributions advance microscopic image analysis through rigorous integration of scanning optimization, distortion correction, and deep learning segmentation. By synthesizing trajectory planning with intelligent algorithms and attention mechanisms, this work establishes practical frameworks for automated high-precision AFM analysis, providing implementation guidance for materials science and biomedical research applications requiring accurate micro-scale morphological characterization in complex imaging environments.