Project #3 Seam Carving

Transcription

1 Project #3 Seam Carving Caroline Hermans For this project, I implemented a Seam Carving algorithm that reduces the size of images without losing important details. Rather than scaling the image, this approach identifies low-importance horizontal and vertical paths through the image ( seams ) and removes them altogether.

2 Algorithm First, through user input, I determine whether to resize the image vertically or horizontally and by how much. Next, I run the image through a gradient function which determines the high and low energy areas of the image by analyzing change between adjacent pixels. The high energy areas correspond to high amounts of information, whereas the low energy areas generally contain less important content. Original lookout.jpg Energy representation of lookout.jpg From there, I find the row or column in the image with the lowest energy value. This corresponds to the row or column of the image that contains the least information. Once I find the lowest energy row or column, I use that information to find the lowest energy value seam through the energy image. Starting from the first pixel in the lowest energy row, I determine which adjacent pixel to the previous pixel has the lowest energy value, store that information, and move on. Using this method, I find a seam, or continuous path, that contains very little important information about the image. Original beach.jpg Low-energy seam through beach.jpg

3 Once a seam has been found, I remove it. I start the removal process with an empty image with one less row or column than the original. Then, I fill in the image row-by-row, removing the one unwanted pixel each time. This entire process repeats as many times as necessary, calculating the lowestenergy seam and removing it until the image is resized to the desired height or width. Examples I performed this process on six example images, resizing them in both the horizontal and vertical directions by 100, 200, and 300 pixels. Original beach.jpg 100px horizontal 200px vertical 200px horizontal 300px vertical 300px horizontal

4 Original fireplace.jpg 300px horizontal 300px vertical

5 300px vertical 200px vertical Original car.jpg Original car.jpg 100px horizontal 200px horizontal 100px

6 Original cmulogo.jpg 100px horizontal 200px horizontal 300px 300px vertical 200px vertical Original lookout.jpg 200px vertical 200px horizontal 300px horizontal

7 Original cmu.jpg 100px horizontal 100px vertical 200px horizontal 200px vertical 300px vertical 300px horizontal

8 Less-Than Perfect Results Squashed Hammerschlag Hall looks very amusing The algorithm actually removed all of Hammerschlag Hall in favor of the trees Here, even though the image is resized, it distorts the car a lot. At this point, it would definitely be better to resize the image by traditional methods.

9 Bells & Whistles Energy Functions I tried two different energy functions to test which one works better under which circumstances. One of the energy functions focuses on the magnitude of the image gradient at each pixel, while one focuses on the direction of the image gradient. Gradient magnitude energy function Gradient direction energy function It became clear that the magnitude energy function is best in most cases, because the direction energy function turns out so noisy that it s difficult to tell which areas are low energy and which are high energy. The gradient direction energy function works best on images with less directional change between adjacent pixels in low-information areas. To test these two different functions, I ran them both on the lookout.jpg and the fireplace.jpg examples and compared the results. Gradient magnitude seam carving Gradient direction seam carving

10 Gradient magnitude seam carving Gradient direction seam carving The result of the direction seam carving on lookout.jpg is inferior to the gradient magnitude seam carving: the girl s head is completely split in half. This is because directional changes are happening all over the image, so it s difficult to see what information is important using this metric. The fireplace.jpg example looks a bit better, but still, the candlestick is choppily cut in half and the whole image looks far less smooth. Comparison to Manual Photo Editing Here, I ve attempted to make a better version of my algorithm s results using Photoshop. The top image is the automatically generated one, and the bottom one was done in Photoshop. I actually made the girl figure smaller to give her more presence in the photo, and I filled in the background behind her with blue waves.