Learn how receptive fields help [CNNs](https://www.ultralytics.com/glossary/convolutional-neural-network-cnn) see context. Explore why [YOLO26](https://docs.ultralytics.com/models/yolo26/) optimizes this for superior object detection.
In the domain of computer vision (CV) and deep learning, the receptive field refers to the specific region of an input image that a particular neuron in a neural network (NN) "sees" or analyzes. Conceptually, it functions similarly to the field of view of a human eye or a camera lens. It determines how much spatial context a model can perceive at any given layer. As data progresses through a Convolutional Neural Network (CNN), the receptive field typically expands, allowing the system to transition from identifying tiny, local details—like edges or corners—to understanding complex, global structures like entire objects or scenes.
感受野的大小和深度由网络架构决定。 在初始层中,神经元 通常具有较小的感受野,聚焦于微小像素簇以捕捉精细纹理。随着 网络深度增加,诸如 池化层和 步长卷积等操作会有效 对特征图进行下采样。该过程 使后续神经元能够整合来自原始输入更大区域的信息。
现代架构(Ultralytics )都经过精心设计以平衡这些感知区域。若感知区域过窄,模型可能因无法捕捉完整形状而无法识别大型物体。 反之,若受容野过度扩展却未保持分辨率,模型则可能遗漏微小物体。为解决此问题,工程师常采用膨胀卷积(又称空洞卷积)技术,在不降低空间分辨率的前提下扩展受容野——这项技术对语义分割等高精度任务至关重要。
优化感受野对各类人工智能解决方案的成功至关重要。
要全面理解网络设计,区分感受野与类似术语很有帮助:
State-of-the-art models like the newer YOLO26 utilize Feature Pyramid Networks (FPN) to maintain effective receptive fields for objects of all sizes. The following example shows how to load a model and perform object detection, leveraging these internal architectural optimizations automatically. Users looking to train their own models with optimized architectures can utilize the Ultralytics Platform for seamless dataset management and cloud training.
from ultralytics import YOLO
# Load the latest YOLO26 model with optimized multi-scale receptive fields
model = YOLO("yolo26n.pt")
# Run inference; the model aggregates features from various receptive field sizes
results = model("https://ultralytics.com/images/bus.jpg")
# Display the results, detecting both large (bus) and small (person) objects
results[0].show()