• Proposed the first joint denoising and demosaicing framework for color-polarized imaging, addressing both degradations within a shared feature space.

• Designed an encoder-based feature fusion mechanism to enhance reconstruction quality.

• Built a real-world paired polarized image dataset to support model training and evaluation.

• Outperformed existing methods across multiple metrics.

• Design a hardware-based tone mapping algorithm to achieve LDR display of Ultra-HDR images which have dynamic range close to the human eye (130-140dB)

• Implement an end-to-end learnable simplified ISP algorithm that utilizes a small neural network to estimate the global tone mapping parameters for an image, enabling real-time estimation on a 24-bit video stream

• Incorporate end-to-end differentiability in the design of the ISP algorithm to guarantee future expandability. For example, the ISP’s output image can be adapted for target detection tasks, enabling joint optimization between the ISP and the corresponding task, thereby enhancing the output metrics

• Implement GPU-based RGB-IR sensor image ISP

• Design reflection removal algorithm based on the low reflection property of glass to infrared light, use infrared band information to remove the reflection of the visible light band and provide better multi-spectral images.

• Utilize guided filter to accelerate the feature extraction and fusion from RGB-IR image

• Designed and implemented a full-stack simulation pipeline modeling the camera imaging process from optics to sensor.

• Applied spatially varying Point Spread Function (PSF) kernels to simulate lens-induced degradation across the image.

• Built a sensor noise modeling framework to generate realistic pixel-level noise maps.

• Delivered an extensible and modular pipeline that can be adapted to different sensors/lenses, providing high-fidelity synthetic raw data for algorithm development and validation.

• Built a modular CPU/GPU parallel operator library including image operators (grayscale, blur, Sobel), matrix multiplication, and DNN layers (convolution, fully connected) with backpropagation support.

• Optimized CPU kernels using MPI, Pthreads, OpenMP, and developed high-performance GPU kernels using CUDA and Triton.

• Triton/CUDA implementations of image operators and matrix multiplication outperformed baseline PyTorch eager implementations under equivalent settings.

• Implemented a high dynamic range (HDR) video processing pipeline based on Google's HDR+ paper, utilizing multi-frame short-exposure RAW synthesis.

• Integrated HDR+ ISP modules including multi-frame alignment and fusion, filtering for noise reduction, and local tone mapping.

• Developed a graphical user interface (GUI) application using Python and PyQt5 framework for user-friendly operation, supporting RAW image sequence selection and processing, with FFmpeg integration for final HDR video synthesis.