Comic-style remote sensing object detection using a hybrid-domain network with multiscale structural improvement

Comic-style remote sensing (CSR) object detection aims to recognize aerial targets from images whose realistic textures have been transformed into line-dominant illustrated representations. CSR images exhibit stronger confusion between decorative textures and target shapes, as well as more severe structural abstraction than conventional optical remote sensing images. Repetitive point patterns, sketch lines, and illustrated shadows may generate numerous false responses, whereas small aircraft, ships, vehicles, and buildings may be represented by only a few fragmented contours. Existing CNNs tend to smooth these weak structural cues, while Transformer-based detectors require higher computational cost and may still mix style-generated lines with target boundaries. This paper proposes HDNet, a hybrid-domain network that jointly models spatial structure and frequency response for CSR object detection. A multiscale dilated contrast module improves local target perception under different receptive fields. A dynamic high-pass filtering module estimates sample-dependent spectral energy, suppresses redundant low-frequency regions and unsupported artistic line responses, and retains contour frequencies that agree with the spatial prediction prior. The two branches are combined by lightweight structural modulation, and a multiscale position-sensitive objective stabilizes small-target localization. Experiments on CSR benchmarks derived from DOTA, HRSC2016, and xView show that HDNet provides an effective balance among detection accuracy, false-alarm suppression, and computational efficiency. Additional analyses clarify annotation consistency, baseline fine-tuning, spectral-threshold sensitivity, loss-weight selection, and reproducibility.