# Is there a metric for the similarity of two image filters?

## Definitions

An image filter is a matrix $m \in \mathbb{R}^{k_1 \times k_2 \times k_3}$ which gets applied to an image $I \in \mathbb{R}^{l_1 \times l_2 \times l_3}$ as a discrete convolution

$$I'(n_1, n_2, n_3) = \sum_{i=0}^{k_1} \sum_{j=0}^{k_2} \sum_{k=0}^{k_3} I[n_1-i - \lfloor \frac{k_1}{2} \rfloor, n_2 - j - \lfloor \frac{k_2}{2} \rfloor, n_3 - k - \lfloor \frac{k_3}{2} \rfloor] \cdot m[i, j, k]$$

There are some well-known filters like Laplace filters, Prewitt filters, ... (see my interactive example)

For example, for an RGB image $k_3 = 3$ and $k_1, k_2$ are width and height.

## Question

Is there a metric to compare the similarity of image filters?

## Context

Convolutional Neural Networks (CNNs) learn image filters. As they are randomly initialized, the filters they learn are different each time you train them. I am interested in quantifying those differences.

I could, of course, use any metric for elements of $\mathbb{R}^{k_1 \times k_2 \times k_3}$. However, consider the filters

\begin{align} m_1 &= \begin{pmatrix}-1&0&1\\-1&0&1\\-1&0&1\end{pmatrix}\\ m_2 &= \begin{pmatrix}1&0&-1\\1&0&-1\\1&0&-1\end{pmatrix}\\ m_3 &= \begin{pmatrix}-0.9&0.1&1\\-0.9&0.1&1\\-0.9&0.1&1\end{pmatrix}\\ \end{align}

For the image

$m_1$ produces

and $m_2$ produces

You can see a difference, but much less than for the result of $m_3$:

This is probably not captured by most metrics. Another idea was to apply the metrics to the processed images on a given dataset, but this would make the results depend on the dataset and be computationally very intensive.

(In case you want to try image filters yourself with Python: https://gist.github.com/MartinThoma/f51a1044c4abc6c7b81915ef96b7cfbd)

The ‘k-translation correlation’ is probably a good candidate for what you are looking for. It measures the maximum correlation between a pair of two filters $\mathbf{W_i}$ and $\mathbf{W_j}$ achieved by translating one filter up to k steps along any spatial dimension and then selecting the maximum thereof:

$$\rho_k(\mathbf{W_i,W_j})=\max_{(x,y)\in \{-k,...,k\}^2\setminus(0,0)} \frac{\langle\mathbf{W_i}, T(\mathbf{W_j}, x,y)\rangle_f}{\left \| \mathbf{W_i}\right \|_2 \left \| \mathbf{W_j}\right \|_2}\,,$$

where $T(\cdot, x,y)$ refers to the translation of its first operand by $(x,y)$ and $\langle\cdot,\cdot\rangle_f$ denotes the flattened inner product of the two filters (the second of which is translated). Note that both filters are reshaped to column vectors to perform the inner product. For more details refer to Doubly Convolutional Networks (Zhai, Cheng, Lu, Zhang, in Proceedings of 30th Conference on Neural Information Processing Systems (NIPS 2016)).

• Could you please elaborate on how to calculate the flattened inner product of two filters (e.g. $W_i, W_j \in \mathbb{R}^{3 \times 3 \times 5}$)? Commented Nov 11, 2016 at 12:06
• To my best knowledge it is nothing other than a simple reshaping operation. So in the given example, you would reshape a 3D filter of size 3x3x5 to a column vector of size 1x45. Commented Nov 11, 2016 at 12:51
In particular, given an image $I$ and two filters $m_1,m_2$, you can compute the results of applying each filter to $I$, namely, $m_1 \otimes I$ and $m_2 \otimes I$. Then, you can compare the similarity of those two results, say, by computing the correlation between them (say, Pearson correlation coefficient). Denote this by $c(m_1 \otimes I, m_2 \otimes I)$.
Now given a training set $\mathcal{I}=\{I_1,\dots,I_n\}$ of images, you can compute the average value of the correlation coefficient across all images in the training set, and use that as your distance metric:
$$d(m_1,m_2) = \frac{1}{n} \sum_{I \in \mathcal{I}} c(m_1 \otimes I, m_2 \otimes I).$$
You rejected this approach out of the concern that it would make the results dependent on the dataset, but I think that's addressed by choosing a training set that matches the distribution of images you intend to use the filter on. In other words, if you intend to use the filter on images produced by some statistical process, then choose the images in the training set by the same procedure. If you do that, then that concern should evaporate. In particular, thanks to the central limit theorem, if the training set is large enough, $d(m_1,m_2)$ converges to a single "true" value as $n \to \infty$ (as the size of the training set increases).