Abstract: Mold spoilage in stored grain poses a serious threat to food security and grain quality. Although conventional physicochemical assays remain the gold standard for mold assessment, their low efficiency and destructive nature hinder large-scale, timely monitoring. Rapid, non-destructive techniques, such as machine vision, hyperspectral imaging, and electronic noses, can capture complementary cues associated with mold development, including surface appearance, compositional changes, and volatile profiles. While traditional machine-learning approaches have advanced the analysis of such data, their reliance on manual feature engineering limits their ability to capture subtle mold-induced variations, often resulting in suboptimal discrimination. In contrast, deep learning enables end-to-end representation learning with hierarchical feature abstraction, thereby improving detection performance. This review summarizes recent deep learning-based studies on stored-grain mold detection, emphasizing multi-source data acquisition and preprocessing for imaging, spectral, and environmental signals. We compare representative architectures, including convolutional neural networks, long short-term memory networks, Transformers, and hybrid models, discussing their characteristics and application performance. Remaining challenges include dependence on large, high-quality datasets, limited interpretability, and poor generalization in complex storage environments. Future efforts should focus on multimodal data fusion and knowledge-enhanced methods to support more reliable intelligent monitoring.