Enhancing Small-Object Segmentation with Meta-Ensemble Architectures

In applications such as autonomous driving and medical imaging, semantic segmentation of small objects remains a key challenge as traditional models often fail to capture finegrained details and precise boundaries. To address this problem, we propose a novel Meta-Ensemble Architecture, which fuses outputs from ResNet50-based U-Net, UNet++, and DeepLabV3 via a meta-model with depthwise separable convolutions and CBAM to refine small-object segmentation by leveraging adaptive refinement. We first systematically optimize the Decoder as well as the Encoder, identifying JaccardLoss as the optimal loss function and ResNet50 as the most efficient backbone for hierarchical feature extraction. We then constructed a stacking-based model that integrates three models (U-Net, UNet++, and DeepLabV3) with ResNet50 as the Encoder. The meta-model utilizes fusion blocks with deeply separable convolutions and the CBAM attention module to dynamically refine segmentation predictions, focusing on the spatial location and discriminative features of small targets. After evaluation on the CamVid dataset, our proposed framework achieves an overall mIoU of 0.631, which is 7.6% higher than the strongest single model (ResNet50 + UNet++). More importantly, the mIoU for small targets improves from 0.319 to 0.433, showing strong performance in recognizing challenging small targets and complex boundaries. This work provides an effective adaptive solution for high-precision small object segmentation.