<![CDATA[Automatic comic element detection is a critical prerequisite for manga digital analysis applications such as indexing, translation, and accessibility enhancement. Manga109 is a widely used benchmark for this task, yet the rapid progress of real-time object detectors has not been matched by a systematic comparison among state-of-the-art models in this domain, particularly between YOLO-family detectors employing Small-Target-Aware Label assignment (STAL) and transformer-based detectors such as RT-DETR. This study benchmarks YOLO26 (nano, small, and medium variants) against RT-DETR-L for detecting four comic element classes (panel, character, text, and face) on Manga109. All models were trained for 100 epochs at 1024×1024 pixels and evaluated using mAP and per-size AP following COCO conventions. YOLO26m achieves the best performance (mAP@0.5:0.95 of 0.7471), while RT-DETR-L obtains the lowest (0.7001) despite the largest parameter count and slowest inference. For small text detection, YOLO26m outperforms RT-DETR-L by 42.7% relatively, and the AP gap across object sizes narrows monotonically as model size grows, supporting the STAL design claim. RT-DETR-L also exhibits significant training instability. The main contribution is a systematic, multi-dimensional benchmark of YOLO26 against RT-DETR-L on Manga109 that evidence STAL effectiveness for small-text detection and offers practical guidance for selecting real-time detectors in manga image analysis.]]>
Copyrights © 2026
