Boosting feature representation via distinctive-consensus fusion: a plug-and-play module for visual recognition

Accurate visual recognition in complex environments—specifically road and building extraction from remote sensing imagery and traffic sign detection in street views—demands architectures that effectively resolve the inherent tension between fine-grained local details and global contextual dependencies. While Convolutional Neural Networks (CNNs) excel at local feature extraction, they often struggle to model long-range relationships. Conversely, existing fusion mechanisms frequently suffer from feature redundancy, excessively focusing on overlapping regions while neglecting complementary disparities, which leads to blurred segmentation boundaries and missed small targets. To address these challenges, we propose the Distinctive-Consensus Fusion Module (DCFM), a universal, plug-and-play unit designed to enhance feature representation. Its core innovation lies in a dialectical fusion strategy employing a Dynamic Weight Generator to guide two parallel mechanisms: Distinctive Feature Attention (DFA), which isolates discriminative patterns to sharpen boundaries and details, and Semantic Consensus Attention (SCA), which aggregates contextual commonalities to ensure global consistency. Furthermore, we incorporate the Derf mechanism to replace standard Layer Normalization, effectively mitigating the over-smoothing problem and enhancing generalization via implicit regularization. Extensive experiments demonstrate DCFM's versatility. Whether replacing skip connections in U-Net for segmentation or embedded into the YOLOv8 framework for object detection, our method achieves superior performance, validating its capability to harmonize distinctive details with semantic consensus across diverse application domains.