Object detection models often face significant performance limitations. These challenges include severe sample imbalance, background interference, and target occlusion. Such issues are particularly prevalent in complex industrial and medical imaging domains. While existing solutions typically focus on data resampling or loss function re-weighting to handle imbalance, a fundamental bottleneck within the network architecture itself is often overlooked. Traditional feature fusion necks, such as FPN and PANet, rely on static convolutions that inevitably become biased towards the majority class during training, leading to the marginalization or loss of minority-class features. To address this critical issue at the feature-fusion level, we propose the Semantic-Aware Fusion Neck (SAF-Neck), which replaces the static fusion paradigm with a dynamic, inputadaptive mechanism. By generating content-aware convolutional kernels for each input, SAF-Neck adaptively enhances the discriminative features of minority-class samples, preventing them from being suppressed by the majority class. We integrate this core innovation into a synergistic architecture with a Lightweight Probabilistic Spatial Attention-HGNetv2(LPSA-HGNetv2) and an imbalance-robust loss function , forming a comprehensive "front-end feature enhancement and back-end optimization" pipeline. We validate our model, SAF-YOLOv11, on a highly challenging industrial task of coal and gangue classification, characterized by a severe class imbalance ratio of up to 1 : 22. Experimental results show that our model achieves a 90.4% F1-score with a computational load of only 5.7 GFLOPs, outperforming the baseline by 4.1% in F1-score while being 13.6% more computationally efficient.
Yang et al. (Sun,) studied this question.