To address the issues such as difficulty in defect representation

insufficient semantic information

and ambiguous class boundaries in corrosion detection of distribution-line fittings

a domain knowledge-guided vision-language detection method is proposed. First

a low-contrast feature enhancement module is designed to enhance textures

boundaries

and key channel responses of low-contrast corrosion regions through multi-scale directional extraction

boundary-guided modulation

and channel adaptive recalibration. Second

a knowledge-guided representation module is constructed to transform distribution-line domain knowledge into textual prompts and to dynamically optimize text embeddings with a vision-conditioned prompt adapter

improving the model’s semantic understanding of corrosion patterns and fitting states. Finally

a cross-modal semantic alignment strategy is introduced to constrain visual-semantic consistency through region-text contrastive classification loss and semantic alignment regularization

enhancing discrimination of similar categories. The experimental results show that the proposed method achieves 87.1% mean average precision (mAP50) on a self-built defect dataset of distribution-line fittings. In open-semantic detection and zero-shot generalization experiments

the mAP50 values reach 81.2% and 44.2%

respectively

with an inference speed of 34.9 frames per second. The proposed method can achieve a balance among detection accuracy

open-category generalization and inference efficiency

providing a technical support for defect detection of distribution-line fittings.