The title is a bit ambitious. More specifically this is a cheat sheet for some keras merging layers, (untrained) convolutional layers, and activation functions with a comparison of luminance, RGB, HSV, and YCbCR color spaces.

After experimenting with concatenate layers, I looked at the other merging layers and decided I needed visual examples of the operations. In these cases, the grayscale/luminance flavors would be the most relevant to machine learning where you're typically working with single-channel feature maps (derived from color images).


I continued with samples of pooling layers and trainable layers with default weights. The latter provided helpful visualizations of activation functions applied to real images.

Add

Add layer applied to a single-channel image.

Add in RGB.

Add in HSV gives you a hue shift.

Adding the Cb and Cr channels gives this color space even more of a hue shift.

Subtract

Subtraction layer applied to a single-channel image.

Subtraction in RGB.

Subtraction in HSV.

Subtracting YCbCr is pretty deformative.

Multiply

Multiply, I guess, makes things darker by the amount of darkness being multiplied (0.0-1.0 values).

RGB multiply looks similar.

In HSV, the multiplication is applied less to brightness and more to saturation.

Likewise YCbCr shifts green.

Average

Average in luminance is pretty straightforward.

Average in RGB also makes sense.

Average in HSV sometimes sees a hue shift.

Average YCbCr works like RGB.

Maximum

Max in monochrome selects the brighter pixel.

Same in RGB.

It's not as straightforward in HSV where hue and saturation impact which pixel value is used.

Max for YCbCr likewise biases toward purple (red and blue) pixels.

Minimum

Minimum, of course, selects the darker pixels.

Same with RGB.

In HSV, minimum looks for dark, desaturated pixels with hues happening to be near zero.

YCbCr looks for dark, greenish pixels.

Most convolutional neural networks use max pooling to reduce dimensionality. A maxpool layer selects the hottest pixel from a grid (typically 2x2) and uses that value. It's useful for detecting patterns while ignoring pixel-to-pixel noise. Average pooling is another approach that is just as it sounds. I ran 2x2 pooling and then resized the output back up to match the input.

Max pooling

In monochrome images you can see the dark details disappear as pooling selects the brightest pixels.

RGB behaves similar to luminance.

HSV makes the occasional weird selection based on hue and saturation.

Much like with maximum and minimum from the previous section, maxpooling on YCbCr biases toward the purplest pixel.

Average pooling

The jaggies (square artifacts) are less obvious in average pooling.

Edges in RGB look more like antialiasing, flat areas look blurred.

HSV again shows some occasional hue shift.

Like with averaging two images, average pooling a single YCbCr image looks just like RGB.

A couple notes for the trainable layers like Dense:

• I didn't train them, just used glorot_uniform as an initializer. So every run would be different and this wouldn't be reflective of trained layers.
• I stuck to monochrome for simplicity; single-channel input, single layer, single-channel output.

The ReLu looks pretty close to identical. I may not understand the layer, but expected that each output would be fully connected to the inputs. Hmm.

Sigmoid looks like it inverts the input.

Softplus isn't too fond of the dark parts of the panda.

Tanh seems to have more or less just darkened the input.

Not really much to observe here except that the dense nodes seem wired to (or heavily weighted by) their positional input pixel.

Update:

This appears to be the case. To get a fully-connected dense layer you need to flatten before and reshape after. This uses a lot of params though.


As with dense, these runs used kernels with default values.

One layer

One conv2d layer, kernel size 3, linear activation.

One conv2d layer, kernel size 3, ReLu activation.

One conv2d layer, kernel size 3, sigmoid activation.

One conv2d layer, kernel size 3, softplus activation.

One conv2d layer, kernel size 3, tanh activation.

This is far more interesting than the dense layers. ReLu seems very good at finding edges/shapes while tanh pushed everything to black and white. What about a larger kernel size?

One conv2d layer, kernel size 7, ReLu activation.

One conv2d layer, kernel size 7, sigmoid activation.

One conv2d layer, kernel size 7, softplus activation.

One conv2d layer, kernel size 7, tanh activation.

Two layers

Two conv2d layers, kernel size 3, ReLu activation for both.

Two conv2d layers, kernel size 3, ReLu activation and tanh activation.

Two conv2d layers, kernel size 3, tanh activation then ReLu activation.

Two conv2d layers, kernel size 3, tanh activation for both.

Transpose

Transpose convolution is sometimes used for image generation or upscaling. Using kernel size and striding, this layer (once trained) projects its input onto a (often) larger feature map.

Conv2dTranspose, kernel size 2, strides 2, ReLu activation.

Conv2dTranspose, kernel size 2, strides 2, sigmoid activation.

Conv2dTranspose, kernel size 2, strides 4, tanh activation.

Conv2dTranspose, kernel size 4, strides 2, ReLu activation.

Conv2dTranspose, kernel size 4, strides 2, tanh activation.

Conv2dTranspose, kernel size 8, strides 2, ReLu activation.

Conv2dTranspose, kernel size 8, strides 2, tanh activation.

Related / internal

Related / external