通道、特征图与感受野

A color photo is not one grid but three: red, green, and blue. Each is a channel, so the image has shape 3 by height by width.

A conv kernel covers every input channel at once. For RGB, a 3x3 kernel is really 3x3x3 and sums across all three.

No matter how many input channels there are, a single kernel produces just one output grid: one feature map.

A layer learns many kernels. Its out_channels count is exactly how many feature maps it stacks into the output.

That layer takes 3 input channels and outputs 32 feature maps. Each map is one learned kernel scanning the RGB image.

The 32 feature maps feed the next conv layer as its input channels. So depth grows as layers stack on top of each other.

A receptive field is the region of the original image that influences one output value. Early layers see only a tiny patch.

Stack more conv layers and each value depends on a wider area. Deep neurons have a large receptive field spanning much of the image. 🌐

With small receptive fields, early layers detect simple things: edges, corners, and color blobs. They are the network's basic features.

Deeper layers combine those into richer patterns: textures, shapes, and eventually whole objects like an eye or a wheel. This is the feature hierarchy.

As you go deeper, maps usually get smaller but gain more channels. The network swaps spatial detail for semantic meaning.

Let us check how channels relate to feature maps.

You learned that kernels span all input channels, out_channels stacks feature maps, and growing receptive fields turn edges into whole objects. 🎉