DCT-Based Transparent Object Segmentation Network with Cooperative CNN–Transformer Branches

Transparent object segmentation is a critical computer vision task for industrial and mobile robotic perception, which is widely deployed in intelligent cleaning robots and unmanned aerial vehicles (UAVs) equipped with visual image sensors. However, the perception performance of mainstream image sensors is severely hindered by the unique optical characteristics of transparent objects, including light transmission, complex reflective noise and indistinct object textures, resulting in inaccurate segmentation results in actual robot sensing scenarios. Precise extraction of transparent object boundaries and suppression of background semantic interference are essential for high-quality segmentation. Existing transparent object segmentation methods either focus on semantic or boundary information independently, or they extract boundary features merely from shallow layers during joint modeling, which always causes incomplete boundary feature extraction. Considering that transparent object boundaries correspond to high-frequency features while low-frequency components carry dominant semantic information, we develop a DCT-based segmentation network with cooperative CNN–Transformer branches. DCT decomposes input images into high- and low-frequency components, where CNN and Transformer branches learn frequency-specific features separately. We embed an interactive information module (IIM) in dual encoders to realize bidirectional cross-frequency feature interaction, strengthen weak boundaries, eliminate incomplete target regions caused by reflection noise, and reduce false edges. In decoders, the hybrid adaptive calibration module (HACM) fuses dual-branch semantic and boundary features, alleviates feature distribution discrepancies, and optimizes object positional information. A cascaded head further achieves effective coarse-to-fine multi-scale feature aggregation to adapt to multi-distance sensing scenarios for robots. Our experimental results on five public datasets verify that our method outperforms current mainstream state-of-the-art segmentation methods, effectively compensating for the limitation of traditional image sensors in transparent object recognition and offering a promising development in the environmental perception robustness of cleaning robots and UAVs.