Household Gas Facility Detection Using RT-DETR with Morphology-Aware Feature Enhancement and Short-Axis-Aware Regression

Detecting household gas facility components in inspection images is challenging because of background clutter, partial occlusion, and substantial variation in object size and shape. This study presents an enhanced RT-DETR-R18 detector combining morphology-aware feature enhancement, local–global interaction, and geometry-sensitive regression. Gradient-Guided Morphology Enhancement (GGME) uses gradient-guided fusion of standard and directional dynamic-sampling branches to represent regular and elongated structures. A Bidirectional Local–Global Interaction Encoder (BLGIE) adaptively combines local structural cues with global context. Short-Axis-Aware Projection Regression (SAPR) supplements the original regression losses with weighted one-dimensional projection alignment and shape regularization. Experiments use 21,813 images containing 47,169 annotated instances across eight categories, with scene-level separation between training and validation sets. Relative to RT-DETR-R18, the proposed detector increases COCO-style mean average precision (mAP) from 67.67% to 71.08%, AP at an intersection-over-union threshold of 0.75 (AP75) from 74.49% to 79.87%, and small-object AP from 40.60% to 49.28%. With 22.3 million parameters and 63.7 GFLOPs, it achieves 72.3 frames per second on an NVIDIA GeForce RTX 4060 Laptop GPU at 640 × 640 resolution using batch-1 FP32 inference. These results indicate improved detection on the evaluated validation set, particularly for stricter localization and small objects.