DeepLabV3+ with MobileNetV3 backbone enhancement and multi-level fusion decoder for pepper field weed segmentation

Accurate crop-weed segmentation is critical for precision weed management in pepper fields. However, reliable pixel-level discrimination remains challenging because pepper plants and weeds often share similar green textures under conditions of occlusion, illumination variation, and scale changes, while weed regions are frequently small, fragmented, and spatially dispersed.

This study proposes an improved DeepLabV3+-based semantic segmentation framework for pepper-weed segmentation. MobileNetV3 was employed as an efficiency-oriented backbone, while Adaptive Activation Fusion and Efficient Channel Attention were integrated to enhance nonlinear feature representation and channel-wise recalibration. A customized atrous spatial pyramid pooling module was designed to capture multiscale contextual information, and an OS = 16 hierarchical decoder was developed to fuse high-, mid-, and low-level features with residual refinement for improved boundary recovery. During training, focal cross-entropy loss and Dice loss were combined with weighted sampling, class-aware cropping, and data augmentation to strengthen the learning of difficult weed regions and improve robustness under complex field conditions.

On the independent pepper-field weed test set, the proposed model achieved an mIoU of 95.87%, recall of 97.98%, a boundary F1-score of 95.88%, a boundary IoU of 92.13%, and an inference speed of 46.86 FPS.

The results demonstrate that the proposed framework improves region-level segmentation accuracy, boundary quality, and inference efficiency. These advantages indicate its potential for practical precision weeding applications in pepper fields.