I computed the structural similarity index (SSIM) value between a ground truth mask and its corresponding predicted mask in an image segmentation task using a UNet model. Both the ground truth and predicted masks are of 1024 x 1024 resolution. I got an SSIM value of 0.9544. I resized the ground truth and predicted mask to 512 x 512 using bicubic interpolation and measured its SSIM value to be 0.9444. I repeated the process for 256 x 256, 128 x128, and 64 x 64 image resolutions and found the SSIM values as 0.9259, 0.8593, and 0.8376. I observed that the equations for luminance, structure, and contrast components in the SSIM formula appear to be normalized and does not seem to vary with image resolution. My question is for the same pair of ground truth and predicted mask, why the SSIM values keep decreasing with decreasing image resolution?
A possible reason: As the resolution of an image decreases, the local patterns become less detailed, leading to a lower SSIM score. Additionally, when the resolution of an image is decreased, the image is effectively being compressed, which can lead to loss of information and further decrease the SSIM score.
To elaborate, the SSIM index is based on three components: the luminance component, which compares the average pixel intensity of the two images; the contrast component, which compares the standard deviation of the pixel intensities; and the structural component, which compares the similarity of patterns in the two images (https://www.mathworks.com/help/images/ref/ssim.html). When the resolution of an image is decreased, the number of pixels in the image is reduced, which can lead to loss of detail in the image. This loss of detail can result in lower values for the luminance and contrast components of the SSIM index, which in turn can lead to a lower overall SSIM score. In addition to this, the structural component of the SSIM index compares the similarity of patterns in the two images using a windowed function, which is sensitive to the resolution of the image. If the resolution is low, the window function captures less information and thus, the structural component becomes less effective in capturing the similarities between two images.
The structural similarity index (SSIM) value is a measure of the structural similarity between two images. It compares the luminance, contrast, and structure of the images. When you decrease the image resolution, you are essentially reducing the number of pixels in the image. This can lead to a loss of information and detail, which can affect the SSIM value.
When you use bicubic interpolation to resize the image, you are essentially estimating the values of the pixels in the new resolution. These estimated values may not match the original high-resolution values exactly, leading to a decrease in the structural similarity between the two images.
Additionally, as the resolution decreases, the structural similarity will be less and less, as the change in pixel values is more pronounced for smaller images. This can also contribute to the decrease in SSIM values as the resolution decreases.
It's worth noting that SSIM is a relative measure and its value can vary depending on the image content, lighting conditions and the specific dataset. The value of 0.9544 for 1024 x 1024 resolution is considered as a high value but it does not provide a clear interpretation of the quality of the predicted mask compared to the ground truth mask.