A5040918025- CCD and CMOS Imaging Sensors
- Analytical Chemistry and Sensors
- Infrared Target Detection Methodologies
- Image Processing Techniques and Applications
- Silicon Nanostructures and Photoluminescence
- Brain Tumor Detection and Classification
- Medical Image Segmentation Techniques
- Advanced Fluorescence Microscopy Techniques
- Advanced Image and Video Retrieval Techniques
- Advanced Optical Sensing Technologies
- Adversarial Robustness in Machine Learning
- Advanced Neural Network Applications
- Anomaly Detection Techniques and Applications
Seoul National University
2024
OmniVision Technologies (United States)
2020-2021
Samsung (South Korea)
2014
자율 주행 차량이 다양한 가시성 조건에서도 안정적으로 작동하기 위해서 낮 시간에 수집된 이미지뿐만 아니라 야간 이미지와 같은 데이터도 필수적이다. 그러나 밤 데이터는 수집이 데이터에 비해 어렵기 때문에 이를 보완하는 이미지 변환 기술이 필요하다. 본 연구에서는 CycleGAN 과 CBAM 을 결합하여 이미지를 이미지로 변환하는 과정에서 성능을 개선하는 방법을 제안한다. Reduction 비율을 조정하고 LP 및 LSE Pooling 방식을 추가해서 모델의 성능과 계산 효율성을 모두 향상시켰다. 실험 결과 8 + 조합이 기존 모델에 성능이 개선됨을 확인하였으며 주행과 응용 분야에서 대량의 합성 데이터를 신속하게 생성할 수 있어 비용 절감 측면에서도 이점을 제공한다.
According to the trend towards high-resolution CMOS image sensors, pixel sizes are continuously shrinking, and below 1.0μm, now reaching a technological limit meet required SNR performance [1-2]. at low-light conditions, which is key metric, determined by sensitivity crosstalk in pixels. To improve sensitivity, technology has migrated from frontside illumination (FSI) backside illumiation (BSI) as size shrinks down. In BSI technology, it very difficult further increase of near-1.0μm because...
This paper presents a 64 mega-pixel, backside-illuminated, CMOS image sensor using 0.7um pixel pitch with 7.0ke- linear full well capacity (FWC). A switchable conversion gain design was also demonstrated to have high 18.0ke- FWC in 4-Cell binning mode. Several new processes were implemented overcome performance degradation due scaling. As result, this achieves low dark noise of 1.26e- and quantum efficiency, comparable larger products, such as 0.8um.
In our work, we have empirically found that Vision Transformer (ViT) could not extract object-centric features when applied to out-of-distribution (OOD) detection. To make attention, design an additional module employs a cross-attention between class-wise token proxy and feature sequence of input image. For inference suitable structure with multiple proxies, propose score ensemble can be any scoring function. Compared ViT, the proposed scheme achieves outperforming performance by synergizing...