MambaSeed: A lightweight Mamba-based hybrid vision network for tomato leaf disease recognition

Crop diseases threaten agricultural productivity, creating a need for recognition models that combine accuracy, robustness, and computational efficiency. Existing CNN-, Transformer-, and state-space-based approaches can weaken fine-grained lesion edges and textures when emphasizing global context. We therefore propose MambaSeed, a lightweight hybrid network built on MambaVisionMixer and equipped with the Local Continuous Branch Injection mechanism, hereafter LCBI. This mechanism repeatedly introduces local feature compensation during hierarchical modeling, preserving high-frequency cues while maintaining global dependency modeling. On the Tomato, Corn, and Potato datasets, MambaSeed achieved accuracies of 98.97%, 98.06%, and 95.87%, respectively, with 1.13M parameters and 0.240G MACs. In the batch-one FP32 GPU benchmark, its latency was 3.39 ms/image on an NVIDIA RTX 4090. Without a hardware accelerator, the median CPU latency was 25.35 ms/image with one thread and 15.61 ms/image with four threads. Ablation, cross-architecture frequency analysis, and controlled perturbation tests further showed that LCBI strengthens high-frequency feature retention and maintains stable recognition under visual degradation. Together, these results show that MambaSeed provides a compact balance of accuracy, efficiency, and robustness across both accelerator-based and CPU-only inference settings.