Enhanced training data acquisition system for artificial intelligence-enabled camera in smartphones

Abstract

With the rapid advancement of deep learning (DL) within the field of artificial intelligence (AI), computer vision technologies have been increasingly integrated into smartphone cameras. Developing DL-based solutions for AI-enabled smartphone cameras, however, demands large volumes of high-quality training data. To address this challenge, this paper introduces a Dual-Camera Real-Image (DCRI) data acquisition system and demonstrates that pretrained networks can be further enhanced through fine-tuning on the proposed dataset. Specifically, the DCRI system comprises a beam splitter, a smartphone camera, and a high-performance digital single-lens reflex (DSLR) camera. We also propose a post-processing pipeline that aligns and color-corrects the paired images, effectively resolving the alignment difficulties commonly observed in prior methods. Extensive experiments confirm that models trained on the DCRI dataset for deep learning-based image signal processing (DL-ISP) achieve substantial improvements in image detail and noise reduction compared with existing approaches. The proposed dataset is publicly available for download.