Kizil Caves murals, as important carriers of cultural heritage, face severe information damage due to natural erosion and human destruction, urgently requiring efficient retrieval technologies to provide scientific basis for restoration and content reconstruction. In cave mural images, each image corresponds to a single semantic label, but contains multiple target elements that collectively represent the label’s meaning, presenting unique challenges of multi-target coexistence, high inter-class similarity, and background interference. To address these challenges, this paper proposes a Multi-Target Feature Mining Hashing for Fine-Grained Cave Mural Retrieval (MTFM). The method adopts a dual-path architecture to mine discriminative features from multiple targets within each mural image. A Triple Fusion Attention Module (TFAM) generates highly discriminative attention maps by fusing multiple attention mechanisms for multi-target localization; a Hierarchical Area Region Detection module (HARD) employs complementary detection strategies to precisely extract multiple key regions; Vision Mamba networks with semantic enhancement Inter-region Semantic Enrichment Module (ISEM) modules are introduced to model long-range dependencies among multi-target regions and generate effective multi-target hash representations. Experimental results on the self-constructed Kizil Caves dataset demonstrate that this method significantly outperforms existing methods across different hash code lengths, achieving improvements of 7.18% \(\sim \) 13.21%.