This thesis addresses two fundamental challenges commonly encountered in image-based recognition systems: multispectral image registration across different wavelength bands and the presence of invalid regions in stereo disparity maps. To resolve these...
This thesis addresses two fundamental challenges commonly encountered in image-based recognition systems: multispectral image registration across different wavelength bands and the presence of invalid regions in stereo disparity maps. To resolve these issues, each problem is analyzed independently, and practical methodologies for data quality enhancement are proposed.
First, to resolve geometric misalignment between spectral bands in a multi-camera-based multispectral acquisition environment, a registration pipeline based on camera calibration and pixel mapping is established. Since each spectral channel is acquired simultaneously through different camera and filter combinations, reprojection to a reference camera coordinate system is performed to account for parallax and lens distortion. Furthermore, a calibration procedure is applied to mitigate radiometric differences across wavelength bands and spatial non-uniformity caused by lens shading, thereby ensuring relative signal consistency. The proposed multispectral registration and correction process is designed for real-time operation even with low-cost hardware configurations, and its effectiveness in corrosion detection applications is validated using real-world ship hull imaging data.
Second, to address invalid regions in Ground Truth disparity maps generated by stereo matching, a disparity map refinement method based on Maximum A Posteriori estimation is proposed. The proposed method constructs a posterior distribution by modeling the distribution of valid neighboring disparities as a prior and the similarity between left–right image patches as a likelihood term, and estimates disparity values for invalid pixels by maximizing this posterior. Experimental results on both synthetic and real-world datasets confirm that the proposed method effectively compensates for invalid regions while maintaining superior structural consistency compared to conventional interpolation-based and learning-based approaches.
In conclusion, this thesis contributes to enhancing the reliability and utility of image-based recognition pipelines by providing practical data quality improvement methods that analyze multispectral registration and disparity map refinement from their respective physical and probabilistic perspectives.