Implementing a custom convolution operation in Tensoflow/Keras

Question

I want to create a convolutional layer that goes over an array with a window, grabs the highest and lowest values and returns a value inversely proportional to the range.

This is the code I'm using presently, but I don't think the operation is applying the way I expect.

class StandardizedConv2DWithOverride(layers.Conv2D):
    def convolution_op(self, inputs, kernel):
        ranger = np.ptp(kernel) #this finds the range in the kernel
        return tf.nn.conv2d(
            inputs,
            1/ranger, #This is the inversely proportional value
            padding="VALID",
            strides=list(self.strides),
            name=self.__class__.__name__,
        )
    
conv = StandardizedConv2DWithOverride(filters=1,kernel_size=(5,5))
conv.build(input_shape=(5,5,1))
conv.set_weights([np.ones((5,5,1,1)),np.zeros((1))])

This should then ideally go through and highlight the uniform areas of this image

But instead it seems to just like blur it

Is there some way of creating such a convolution layer?