Rice Seed Vitality Detection Based on WGAN and Improved Lightweight EDCA-MobileNet V3

To address the challenges of high model redundancy, skewed data distribution, and insufficient finegrained feature representation in rice seed vitality detection, this study proposes a lightweight detection model incorporating an Efficient Dual-Channel Attention (EDCA) mechanism. First, a dataset comprising five categories of rice seeds was constructed, and data augmentation was performed using Wasserstein Generative Adversarial Networks (WGAN) to mitigate data scarcity and class imbalance. For network architecture optimization, a three-stage improvement strategy was implemented: 1) Redundant modules in MobileNet V3 were eliminated through a hierarchical pruning strategy, reducing parameter size and enhancing inference speed; 2) An EDCA mechanism was designed, integrating multi-dimensional feature compression-reorganization, 1D convolutional dynamic weight allocation, and parallel dual-path fusion of multi-granularity features to improve fine-grained feature extraction; 3) Global average pooling replaced max pooling to minimize information loss, Dropout regularization was introduced to suppress overfitting, and SE Blocks were removed to balance complexity and performance. Comparative experiments with four modelsMobileNet V3, CBA-MobileNet, WGAN-MobileNet V3, and EDCA-MobileNet V3-demonstrated the superior performance of EDCA-MobileNet V3, achieving a 42.6% reduction in parameters (3.1 M total), a Top-1 accuracy of 96.5%, and improvements of 8.3%, 6.2%, and 3.6% over the baseline models, respectively. Cloud deployment exhibited a 28.3% reduction in end-to-end latency, 21.9% lower GPU utilization, and 68.4% compression in peak memory usage, validating the robustness and real-time advantages of the improved EDCA-MobileNet V3 in high-concurrency agricultural IoT scenarios.