To Detect, and Beyond: Integrating Text-Guided Object Detection and Super-Resolution

초록

Object detection in conventional surveillance systems is typically constrained to predefined objects. Furthermore, the low resolution of the video often impairs the identifiability of these detected objects. In this paper, we propose a novel framework that not only detects unseen objects but also enhances the identifiability of detected objects. The proposed framework introduces text guidance into object detection and super-resolution processes, enabling the detection of untrained objects and restoring them to high resolution while preserving their details. Moreover, text-guided object detection and super-resolution are synergistically integrated through the shared text encoding between the two processes. Our evaluation on a real world dataset demonstrates the effectiveness of the proposed framework in minimizing distortion of detected objects and enhancing identifiability by fusing information from different modalities of image and text during restoration. © 2024 IEEE.

키워드