Chen Tianqi, Todo Yuki, Qiu Zhiyu, Hua Yuxiao, Sugiura Hiroki, Tang Zheng
Division of Electrical Engineering and Computer Science, Kanazawa University, Kanazawa 920-1192, Japan.
Faculty of Electrical, Information and Communication Engineering, Kanazawa University, Kanazawa 920-1192, Japan.
Biomimetics (Basel). 2025 May 2;10(5):286. doi: 10.3390/biomimetics10050286.
Motion direction detection is an essential task for both computer vision and neuroscience. Inspired by the biological theory of the human visual system, we proposed a learnable horizontal-cell-based dendritic neuron model (HCdM) that captures motion direction with high efficiency while remaining highly robust. Unlike present deep learning models, which rely on extension of computation and extraction of global features, the HCdM mimics the localized processing of dendritic neurons, enabling efficient motion feature integration. Through synaptic learning that prunes unnecessary parts, our model maintains high accuracy in noised images, particularly against salt-and-pepper noise. Experimental results show that the HCdM reached over 99.5% test accuracy, maintained robust performance under 10% salt-and-pepper noise, and achieved cross-dataset generalization exceeding 80% in certain conditions. Comparisons with state-of-the-art (SOTA) models like vision transformers (ViTs) and convolutional neural networks (CNNs) demonstrate the HCdM's robustness and efficiency. Additionally, in contrast to previous artificial visual systems (AVSs), our findings suggest that lateral geniculate nucleus (LGN) structures, though present in biological vision, may not be essential for motion direction detection. This insight provides a new direction for bio-inspired computational models. Future research will focus on hybridizing the HCdM with SOTA models that perform well on complex visual scenes to enhance its adaptability.
运动方向检测对于计算机视觉和神经科学来说都是一项至关重要的任务。受人类视觉系统生物学理论的启发，我们提出了一种基于水平细胞的可学习树突神经元模型（HCdM），该模型能够高效捕捉运动方向，同时保持高度鲁棒性。与目前依赖于扩展计算和提取全局特征的深度学习模型不同，HCdM模仿了树突神经元的局部处理方式，实现了高效的运动特征整合。通过修剪不必要部分的突触学习，我们的模型在有噪声的图像中保持高精度，特别是对于椒盐噪声。实验结果表明，HCdM的测试准确率超过99.5%，在10%的椒盐噪声下保持鲁棒性能，并且在某些条件下跨数据集泛化率超过80%。与视觉Transformer（ViT）和卷积神经网络（CNN）等当前最先进（SOTA）模型的比较证明了HCdM的鲁棒性和效率。此外，与以前的人工视觉系统（AVS）相比，我们的研究结果表明，外侧膝状体（LGN）结构虽然存在于生物视觉中，但对于运动方向检测可能不是必需的。这一见解为受生物启发的计算模型提供了新的方向。未来的研究将集中于将HCdM与在复杂视觉场景中表现良好的SOTA模型进行融合，以提高其适应性。
