Answers are grounded in this paper and cite their sources.
To address the challenge of simultaneously optimizing model accuracy and computational efficiency in the automated inspection of wheat grain appearance quality, this study proposes a lightweight deep learning model called MobileNet-WDD. Based on the MobileNetV4-small architecture, this model incorporates the SimAM attention mechanism to enhance feature discrimination capabilities and employs Ghost convolutions and the Mish activation function to optimize the network structure, thereby significantly reducing model complexity while maintaining high recognition accuracy. Using six typical categories of wheat grains-diseased spots, insect damage, mold, sprouting, damage, and intact grains-as the research subjects, experimental results show that compared to the baseline model, MobileNet-WDD achieves a 6.1% increase in accuracy, reaching 94.2%; a 30.3% reduction in the number of parameters; computational cost by 27.6%; and inference speed from 155 FPS to 183 FPS, representing a 1.18-fold acceleration. Quantitative analysis confirms that this model achieves high-precision recognition while offering significant advantages in terms of lightweight design and efficient computational performance, providing an efficient and feasible technical solution for real-time non-destructive inspection of wheat grains.
No takes yet. Share an insight, caveat, or question.
Ma et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: