A super-fast Python implementation of Seam carving for content-aware image resizing, and the forward energy function proposed in Improved seam carving for video retargeting.

With seam carving algorithm, the image could be intelligently resized while keeping the important contents undistorted. The carving process could be further guided, so that an object could be removed from the image without apparent artifacts.

Install a stable version from PyPI.

pip install seam-carving

Or install the latest version from GitHub.

pip install git+https://github.com/li-plus/seam-carving.git@master

To resize an image using seam carving:

import numpy as np
from PIL import Image
import seam_carving

src = np.array(Image.open("fig/castle.jpg"))
src_h, src_w, _ = src.shape
dst = seam_carving.resize(
    src,  # source image (rgb or gray)
    size=(src_w - 200, src_h),  # target size
    energy_mode="backward",  # choose from {backward, forward}
    order="width-first",  # choose from {width-first, height-first}
    keep_mask=None,  # object mask to protect from removal
)
Image.fromarray(dst).show()

To remove an object from the image:

src = np.array(Image.open("fig/beach.jpg"))
mask = np.array(Image.open("fig/beach_girl.png").convert("L"))
dst = seam_carving.resize(src, drop_mask=mask)
Image.fromarray(dst).show()

For more examples, please refer to example/demo.ipynb.

Resizing along the x-axis using the original backward energy function:

Reducing the width using backward & forward energy functions:

The image width and height could be changed simultaneously in an optimal seam order, but the order has little effect on the final result. Currently we only support two kinds of seam orders: width-first and height-first. In width-first mode, we remove or insert all vertical seams first, and then the horizontal ones, while height-first is the opposite.

The protected mask is not affected by seam removal and insertion.

Specify an object mask to remove (red) and a mask to protect (green, optional).

We compare the performance of our implementation and other popular Python repos on castle.jpg (600x407). The image is narrowed or widened by 200 pixels using backward energy (BE) or forward energy (FE), respectively. Below is the running time (second) evaluated on a MacBook Pro.

• Avidan, S., & Shamir, A. (2007). Seam carving for content-aware image resizing. In ACM SIGGRAPH 2007 papers (pp. 10-es). [paper] [blog]
• Rubinstein, M., Shamir, A., & Avidan, S. (2008). Improved seam carving for video retargeting. ACM transactions on graphics (TOG), 27(3), 1-9. [paper]
• Das, A. (2019). Improved seam carving with forward energy. [blog]