Jin, MinhyunMinhyunJinChu, MyonglaeMyonglaeChuKara, GökhanGökhanKaraYoo, HyunyoungHyunyoungYooLee, Sang YeonSang YeonLeeSeo, YooraYooraSeoMalinowski, PawelPawelMalinowskiLieberman, ItaiItaiLiebermanGenoe, JanJanGenoeLee, JiwonJiwonLee2026-11-052026-11-0520260018-9383https://imec-publications.be/handle/20.500.12860/60738In-sensor computing, which integrates signal acquisition and information processing within image sensors, is emerging as a promising route toward low-power artificial intelligence hardware. However, its realization has been hindered by the lack of scalable pixel architectures capable of combining high-quality photodetection with computing functionality. In this work, we propose an in-sensor convolution scheme enabled by a thin-film pinned photodiode (TF-PPD) structure. The TF-PPD employs a photogate (PG) that enables dynamic tuning of pixel sensitivity, thereby facilitating the direct implementation of convolution operations within the sensor array. The convolutional neural network (CNN) implemented with the proposed TF-PPD-based convolution technique was experimentally validated through image demonstrations. The proposed technique achieved 98.6% accuracy on the modified National Institute of Standards and Technology (MNIST) and 94.9% on captured handwritten digits, while reducing the computational cost by 39.7%. These results demonstrate that the TF-PPD provides both stable photodetection and embedded computing capability, highlighting its potential as a scalable device-level solution for next-generation intelligent vision sensors.engThin-Film Pinned Photodiode Pixel-Based In-Sensor Convolution ProcessingJournal article10.1109/ted.2026.3696243WOS:001779936300001CMOS IMAGE SENSORNEURAL-NETWORKVISION1557-9646