The misalignment between RGB and thermal images significantly impairs RGB-Thermal semantic segmentation accuracy. Current non-end-to-end methods treat RGB-Thermal registration independently of semantic segmentation, resulting in fusion errors, redundant computations, and poor real-time performance. Semantic segmentation accuracy directly correlates with registration precision: better registration yields more accurate segmentation. Moreover, regions with identical semantic labels, indicating the same object, tend to share similar registration offsets. Based on these correlations, we propose an end-to-end multimodal registration and segmentation method using flexible deformation fields. Our method utilizes a shared encoder for registration and semantic segmentation to reduce redundancy. Unlike traditional non-end-to-end approaches, it directly registers high-level perceptual features, thereby optimizing computational efficiency and real-time performance. Additionally, we employ a flexible deformation field to register RGB-Thermal data, addressing limitations of traditional affine transformations in handling non-coplanar and non-rigid registrations. However, the increased flexibility of deformation fields compared to affine transformations, and the sacrificing of geometric feature preservation, pose training challenges. To overcome this, we introduce a semantic alignment loss function to train the alignment module. This function calculates the semantic segmentation loss between the predictions from registered thermal features and RGB semantic labels. It shortens the gradient backpropagation path, aligning the objectives of registration and segmentation. We validate our end-to-end approach through extensive experiments, achieving significant performance enhancements. On the IR SEG dataset, our end-to-end method achieves state-of-the-art results with a mean Intersection over Union (mIoU) of 61.1% and a mean accuracy (mAcc) of 76.0%.

Single-image depth from defocus is limited by blur-radius ambiguity: two object distances on different sides of the focal plane can produce similar point spread functions (PSFs). We study single-shot dual-focus imaging (DFI) with a polarizer-free liquid crystal (LC) lens, where the ordinary-ray component remains unmodulated while the extraordinary-ray component is refocused. Unlike conventional multi-capture DFD, the proposed system records a dual-focus superposition in one exposure and selects the optical setting before network training using the ambiguity-interval length A and PSF correlation $C_r$. We show that DFI does not reduce ambiguity unconditionally: its benefit depends on the LC-lens power. A deblurring-based depth estimation network with a physics-calibrated Wiener bank translates the selected dual-focus cue into quantitative depth. On model-matched synthetic DFI data with an 8 m focus setting, the proposed configuration reduces RMS error from $0.227\pm 0.001$ m to $0.190\pm 0.001$ m (mean ± s.d. over four independent runs) relative to single-focus imaging. A 116-pair indoor prototype dataset provides feasibility evidence under the tested configuration, but is not used to claim broad physical generalization.