MoEMASeg: An Enhanced DeepLab V3 Combining MobileNet V2 and EMA

Computer vision has become ubiquitous, with artificial intelligence continuously evolving, and deep learning emerging as one of the primary approaches in image processing. While the current DeepLabV3+ based image segmentation algorithm achieves high accuracy, its computational complexity, resource consumption, and intricate feature maps result in substantial processing time, making it unsuitable for real-time applications. In this paper, we propose an improved DeepLabV3 image segmentation algorithm (MoEMASeg). The proposed method employs the lightweight MobileNetV2 as the backbone feature extraction network and incorporates an EMA (Efficient Multi-Scale Attention), which effectively reduces the model’s parameters and computational overhead. Additionally, the EMA module enhances image semantic segmentation performance by facilitating better integration of multi-scale features. We adopt a combined loss function incorporating cross-entropy and Dice coefficient to further improve segmentation accuracy. Experimental validation is conducted on the PASCAL-VOC dataset, evaluating the model across multiple dimensions including accuracy, parameter count, and inference time. The experimental results demonstrate that our improved method significantly reduces the parameter count compared to the original DeepLabV3, while achieving an MIoU of 87.15%. This represents a substantial reduction in algorithmic complexity and execution time while maintaining segmentation quality. The proposed model exhibits robust performance in object contour detection and provides a novel technical solution for real-time lightweight image segmentation applications.