Jeonneung Baek

Objective

Evaluate how super-resolution affects object detection in low-resolution aerial images.

Work Performed

• Downsampled DOTA images to one quarter of their original width and height, then restored them using channel transformers and depthwise separable convolutions.
• Compared models trained with L1 and perceptual loss using Faster R-CNN, RoI Transformer, and ReDet.

Results

Mean mAP across the three detectors was 0.535 for bicubic, 0.712 for SR with L1 loss, and 0.736 for SR with perceptual loss. Perceptual-loss SR improved mean mAP by 0.201 over bicubic in this experiment.

First-author paper at the 2025 Fall Conference of the Korean Institute of Broadcast and Media Engineers. Values are arithmetic means under the reported DOTA experiment conditions.

Objective

Estimate structures and tissue regions in mammography for subsequent processing and improve peripheral visibility and perceived depth.

Work Performed

• Analyzed distribution differences between public and target images; built training data through preprocessing, augmentation, and manual review and correction of model predictions.
• Compared and trained segmentation models, exported them to ONNX, and integrated them into C++/OpenCV software. Addressed false detections and boundary errors through post-processing.

Results

Implemented region estimation models adapted to the target images and AI-assisted enhancement functions. Adjusted the processing to preserve peripheral visibility and perceived depth through subsequent stages.

Industry project in the Image Processing Team at Rayence, using PyTorch, ONNX, C++, and OpenCV.

Objective

Reduce processing time and unnecessary resource use in a mammography image processing engine.

Work Performed

• Measured per-function runtime and reviewed C++ computation structures to identify bottlenecks.
• Reviewed unnecessary allocations and excessive input resolution, then revised the implementation to reduce computational cost.

Results

Improved the runtime and resource use of key image processing functions and incorporated the changes into the software.

Industry project at Rayence: C++ profiling, memory management, and computation optimization.

Objective

Acquire denoising training data that reflects the characteristics of a new dental X-ray sensor.

Work Performed

• Explored acquisition conditions through dose comparison experiments and analysis of sensor images.
• Captured the same objects under different conditions to construct image pairs, organized the data, and supplied it to an external AI developer.

Results

Built a sensor-specific training dataset to support external development of a denoising model.

Industry project at Rayence: sensor characterization, X-ray acquisition, and paired dataset construction.

Objective

Mitigate the loss of image keyword extraction performance caused by reduced resolution.

Work Performed

• Analyzed multi-label classification models and measured the effect of downsampling on keyword extraction.
• Compared classification mAP after bicubic interpolation and restoration with EDSR and MCSR.

Results

In the quarter-resolution restoration experiment, reported mAP was 77.16 for bicubic, 82.83 for EDSR, and 80.93 for MCSR, showing a classification improvement with super-resolution under those conditions.

First-author paper at IPIU 2024: Multi-Label Classification Using Super-Resolution.

Objective

Investigate models and processing pipelines for detecting anomalous events in unattended CCTV video.

Work Performed

• Compared and evaluated unsupervised and weakly supervised video anomaly detection models.
• Explored sliding-window frame processing and super-resolution restoration as ways to improve performance.

Results

Implemented and shared end-to-end CCTV anomaly detection code, covering video input through anomaly prediction.

Includes an informal research internship with PIA Space AI Research · July–August 2024.

Objective

Use transformer representations in super-resolution while maintaining computational efficiency.

Work Performed

• Designed MCT by combining channel-wise transformers with the multi-connection structure of MCSR.
• Applied depthwise separable convolutions and residual connections in RCCT, comparing PSNR and runtime.

Results

MCT improved PSNR over the convolutional comparison model. RCCT improved PSNR and runtime relative to MCT and was also used for aerial image restoration.

First-author conference papers: MCT at the 2024 Summer Conference and RCCT at the 2025 Fall Conference of the Korean Institute of Broadcast and Media Engineers.

Objective

Investigate architectures that reduce computational cost in convolutional super-resolution models.

Work Performed

• In RDSR, applied depthwise separable convolutions and GeLU-based gating within a residual structure based on RLFN.
• In CSPSR, separated feature flows with a Cross Stage Partial architecture and evaluated reconstruction quality and computational efficiency.

Results

Presented RDSR as a first-author paper. The CSPSR work received second prize in the Innovation showcase ×4 Super-Resolution track of the ICIP 2024 challenge.

RDSR: 2024 Summer Conference of the Korean Institute of Broadcast and Media Engineers. The award refers to the specified challenge track.

Objective

Improve feature map coding efficiency while preserving downstream machine vision performance.

Work Performed

• Extracted and mapped features from detection, segmentation, and tracking models, using downsampling and super-resolution reconstruction.
• Compared VVC tool activation settings to identify configurations suited to feature maps.

Results

In the reported comparison, maintained mAP of 0.82 while reducing bitrate by 31.16% and runtime by 35.29% relative to the anchor.

Co-authored paper at IPIU 2023. Figures refer to the specific anchor and proposed configurations in the reported experiment.

Objective

Reduce quality degradation and edge information loss introduced by immersive video compression.

Work Performed

• Applied SwinIR-based neural network post-filters to VVC-coded atlas patches and synthesized views.
• Used Canny edge detection to preserve edge regions during pruning and evaluated quality and BD-rate by sequence.

Results

Analyzed performance differences across filter placement and sequences, contributing post-filtering and edge preservation findings to MPEG-I standardization.

MPEG contributions m64816, m67136, and m67137, plus a co-authored 2024 paper on edge preservation.

Objective

Improve compressed thermal images by selecting restoration models suited to different compression strengths.

Work Performed

• Trained separate NNPFs for HEVC quantization parameter settings and compared PSNR and SSIM.
• Classified image quality with a VGG-based model and adaptively selected the corresponding NNPF.

Results

In the reported mixed-quality experiment, adaptive selection achieved PSNR of 54.68 and SSIM of 0.8627, compared with 54.53 and 0.8559 when always using the high-quality NNPF.

Co-authored paper at the 2024 Summer Conference of the Korean Institute of Broadcast and Media Engineers. Figures apply to the experiment reported in the portfolio.