Enhanced DeepLabV3+ for wheat grain segmentation: Addressing tiny targets, blurred edges, and class imbalance

To address the limitations of traditional semantic segmentation models in wheat grain segmentation tasks, particularly the challenges stemming from tiny target sizes, blurred edges, and class imbalance, this paper proposes an enhanced DeepLabV3+ architecture. Key improvements encompass the adoption of a lightweight MobileNetV2 as the feature extraction backbone, coupled with an adaptive dynamic downsampling strategy to bolster multi-scale feature representation; the incorporation of a Strip Pooling (SP) mechanism into the Atrous Spatial Pyramid Pooling (ASPP) module to capture contextual information of elongated grains via horizontal-vertical dual-path aggregation, thereby mitigating feature loss for slender targets; the integration of a Coordinate Attention (CA) mechanism to enable channel-spatial dynamic weighting and accentuate critical regions; and the construction of a multi-loss optimization framework that integrates Weighted Cross-Entropy (WCE) loss and Dice Similarity Coefficient (Dice) loss to balance class distribution and improve edge segmentation accuracy. Experimental results from a self-constructed wheat grain dataset show that the proposed model achieves a mean Intersection over Union (mIoU) of 86.93% and a mean Pixel Accuracy (mPA) of 95.2%, outperforming the baseline DeepLabV3+ by 9.35% and 5.22%, respectively. These findings suggest that the proposed model can yield more refined segmentation outcomes in wheat grain segmentation tasks. • Proposes an enhanced DeepLabV3+ for wheat kernel segmentation, addressing tiny targets, blurred edges, and class imbalance. • Employs MobileNetV2 + adaptive downsampling, integrates dual - path aggregation (Strip Pooling in ASPP) and Coordinate Attention to enhance feature representation. • Integrates Weighted Cross-Entropy & Dice losses into multi - loss framework, balancing classes and enhancing edge segmentation. • Achieves 86.93% mIoU, 9.35% higher than baseline on self-built dataset.