From Recognition to Reasoning: Edge-Frequency Chain-of-Thought for Remote Sensing Object Detection

Abstract

The main bottleneck of current object detectors does not lie in weak feature representation. It lies in the lack of structured reasoning. Most detectors follow a recognition-style pipeline. They map features directly to detection results. This design ignores structural relations inside visual information. As a result, model stability and robustness degrade in complex scenes. This article proposes edge-frequency chain-of-thought (EF-CoT), a structured reasoning paradigm for object detection. EF-CoT reformulates detection as a stepwise reasoning process. It jointly models domain-specific tokens, spatial priors, and progressive reasoning. The model analyzes visual evidence before making predictions. Based on EF-CoT, this article designs the triple-branch guided frequency alignment and edge-aware network (TGFAENet). TGFAENet builds edge-frequency-aware structured tokens. Token-level reasoning then models the interaction among global structure, local details, and boundary cues. The Gaussian attention spatial pyramid pooling fusion (GASPPF) module is further introduced. GASPPF uses dual-branch spatial pyramid pooling and Gaussian attention (GA) to model continuous spatial priors. This module guides the detector toward potential target regions. By combining structured token modeling, spatial awareness, and chain-style reasoning, the proposed method builds a unified perception–analysis–decision detection framework. Experiments on complex remote sensing scenes show clear performance gains. The gains are more visible on small-object detection. These results support the shift from direct recognition to diagnostic reasoning in object detection.