Qiu Zhiyu, Hua Yuxiao, Chen Tianqi, Todo Yuki, Tang Zheng, Qiu Delai, Chu Chunping
Division of Electrical, Engineering and Computer Science, Graduate School of Natural Science & Technology, Kanazawa University, Kakuma, Kanazawa 920-1192, Japan.
Faculty of Electrical and Computer Engineering, Kanazawa University, Kakuma-Machi, Kanazawa 920-1192, Japan.
Biomimetics (Basel). 2025 May 19;10(5):332. doi: 10.3390/biomimetics10050332.
Motion perception is a fundamental function of biological visual systems, enabling organisms to navigate dynamic environments, detect threats, and track moving objects. Inspired by the mechanisms of biological motion processing, we propose an Unsupervised Artificial Visual System for motion direction detection. Unlike traditional supervised learning approaches, our model employs unsupervised learning to classify local motion direction detection neurons and group those with similar directional preferences to form macroscopic motion direction detection neurons. The activation of these neurons is proportional to the received input, and the neuron with the highest activation determines the macroscopic motion direction of the object. The proposed system consists of two layers: a local motion direction detection layer and an unsupervised global motion direction detection layer. For local motion detection, we adopt the Local Motion Detection Neuron (LMDN) model proposed in our previous work, which detects motion in eight different directions. The outputs of these neurons serve as inputs to the global motion direction detection layer, which employs a Gaussian Mixture Model (GMM) for unsupervised clustering. GMM, a probabilistic clustering method, effectively classifies local motion detection neurons according to their preferred directions, aligning with biological principles of sensory adaptation and probabilistic neural processing. Through repeated exposure to motion stimuli, our model self-organizes to detect macroscopic motion direction without the need for labeled data. Experimental results demonstrate that the GMM-based global motion detection layer successfully classifies motion direction signals, forming structured motion representations akin to biological visual systems. Furthermore, the system achieves motion direction detection accuracy comparable to previous supervised models while offering a more biologically plausible mechanism. This work highlights the potential of unsupervised learning in artificial vision and contributes to the development of adaptive motion perception models inspired by neural computation.
运动感知是生物视觉系统的一项基本功能，使生物体能够在动态环境中导航、检测威胁并跟踪移动物体。受生物运动处理机制的启发，我们提出了一种用于运动方向检测的无监督人工视觉系统。与传统的监督学习方法不同，我们的模型采用无监督学习来对局部运动方向检测神经元进行分类，并将具有相似方向偏好的神经元分组，以形成宏观运动方向检测神经元。这些神经元的激活程度与接收到的输入成正比，激活程度最高的神经元决定物体的宏观运动方向。所提出的系统由两层组成：局部运动方向检测层和无监督全局运动方向检测层。对于局部运动检测，我们采用了我们之前工作中提出的局部运动检测神经元（LMDN）模型，该模型可以检测八个不同方向的运动。这些神经元的输出作为输入提供给全局运动方向检测层，该层采用高斯混合模型（GMM）进行无监督聚类。GMM是一种概率聚类方法，它根据局部运动检测神经元的偏好方向有效地对其进行分类，这与感觉适应和概率神经处理的生物学原理相一致。通过反复接触运动刺激，我们的模型能够自组织以检测宏观运动方向，而无需标记数据。实验结果表明，基于GMM的全局运动检测层成功地对运动方向信号进行了分类，形成了类似于生物视觉系统的结构化运动表征。此外，该系统在实现与先前监督模型相当的运动方向检测准确率的同时，还提供了一种更符合生物学原理的机制。这项工作突出了无监督学习在人工视觉中的潜力，并为受神经计算启发的自适应运动感知模型的发展做出了贡献。
