FFCS-Net: controlled multi-prior interaction via decoupled residual gating for lightweight defect detection

Embedding domain priors can compensate for the limited capacity of lightweight defect detectors, yet existing methods implicitly treat prior stacking as beneficial, overlooking the nonlinear interactions—synergy, perturbation, or conflict—that may arise when heterogeneous inductive biases are injected simultaneously into the same information flow. This paper addresses this gap within the scope of lightweight steel surface defect detection. We propose FFCS-Net, a controlled feature selection framework that decouples three inductive biases—edge gradient, scale preference, and detail preservation—using learnable residual scaling factors for independent intensity adjustment of edge and detail priors, while scale preference is regulated through SE channel gating; the framework is lightweight in parameter count, approximately 2.0M, at a moderate inference-speed trade-off that remains acceptable for offline industrial inspection. Through an eight-run full factorial ablation, we uncover a synergy-disturbance-boundary progression pattern of multi-prior interactions: edge and scale priors exhibit near-linear synergy with a gain ratio of approximately 0.97 on NEU-DET; adding the detail prior induces P-R imbalance on Aero-Engine with a gap of 33.19%, narrowed to 2.79% via scaling factor control; under extreme class imbalance, SE channel gating suppresses minority classes, a failure circumvented by feature concatenation. Experiments on NEU-DET, Aero-Engine, and GC10-DET validate these findings, suggesting that the injection intensity, rather than the mere presence, of inductive biases governs their interaction outcomes in lightweight detector design.