Sensitivity to image noise

To quantitatively check the robustness of our method to image noise, we added multiplicative noise to every pixel's intensity value:

$$I_{noise} = I_{orig} + \sqrt{v}*rand(0,1)$$ (5.1)

We assume $I_{orig}$ between 0 and 1. We were increasing $v$ from 0.01 to 0.3 for trucks and from 0.01 to 0.1 for car images. Truck images have higher resolution, thus the variance $v$ could be bigger.

This experiment is implemented as follows:

• Step 1. Identify learning set $D$ and testing set $T$.
• Step 2. Initialize $V$ with chosen variance: $V=\{0.01, 0.02, 0.04, 0.06, 0.08, 0.1, 0.2, 0.3\}$. Let $v$ is the first member of $V$;
• Step 3. $T_v=\{\}$. For every image $I_{orig}$ in testing set $T$, add multiplicative noise of variance $v$ to every pixel according to the equation 5.1 and construct the new noisy image $I_{noise}$. Add this image $I_{noise}$ to newly created testing set $T_v$.
• Step 4. Do the classification with noisyfied testing images $T_v$ and learning set $D$.
• Step 5. Let $v$ is the next member of $V$.
• Step 6. Goto Step 3.
• Step 7. Plot a 2-D graph for every variance $v$ and corresponding classification rate.