Caire is a content aware image resize library based on Seam Carving for Content-Aware Image Resizing paper.
- An energy map (edge detection) is generated from the provided image.
- The algorithm tries to find the least important parts of the image taking into account the lowest energy values.
- Using a dynamic programming approach the algorithm will generate individual seams across the image from top to down, or from left to right (depending on the horizontal or vertical resizing) and will allocate for each seam a custom value, the least important pixels having the lowest energy cost and the most important ones having the highest cost.
- We traverse the image from the second row to the last row and compute the cumulative minimum energy for all possible connected seams for each entry.
- The minimum energy level is calculated by summing up the current pixel value with the lowest value of the neighboring pixels obtained from the previous row.
- We traverse the image from top to bottom and compute the minimum energy level. For each pixel in a row we compute the energy of the current pixel plus the energy of one of the three possible pixels above it.
- Find the lowest cost seam from the energy matrix starting from the last row and remove it.
- Repeat the process.
| Original image | Energy map | Seams applied |
|---|---|---|
![]() |
![]() |
![]() |
Key features which differentiates this library from the other existing open source solutions:
- GUI progress indicator
- Face detection support to avoid face deformation
- Text detection support to avoid text deformation
- Customizable command line support
- Support for both shrinking or enlarging the image
- Resize image both vertically and horizontally
- Support for multiple output image type (jpg, jpeg, png, bmp, webp)
- Support for
stdinandstdoutpipe commands - Can process whole directories recursively and concurrently
- Support for squaring the image with a single command
- Support for proportional scaling
- Support for protective mask
- Support for removal mask
- GUI debug mode support
First, install Go, set your GOPATH, and make sure $GOPATH/bin is on your PATH.
$ go install github.com/esimov/caire/cmd/caire@latestThe library can also be installed via Homebrew.
$ brew install caire$ caire --helpThe following flags are supported:
| Flag | Type | Description | Default Value |
|---|---|---|---|
in |
- | Input file | |
out |
- | Output file | |
width |
n/a | Desired image width | |
height |
n/a | Desired image height | |
preview |
bool | Show GUI window | true |
debug |
bool | Enable/disable debugging in preview mode | false |
perc |
bool | Reduce image by percentage | false |
square |
bool | Reduce image to square dimensions | false |
blur |
int | Edge detector blur radius | 4 |
sobel |
int | Sobel filter threshold | 2 |
faceDetect |
bool | Activate face detection | false |
textDetect |
bool | Activate text detection | false |
textDetInterval |
int | The number of cycles after the text detector should be re-activated | 1 |
cannyHigh |
float | Canny edge detector high threshold | 300 |
cannyLow |
float | Canny edge detector low threshold | 50 |
gaussian |
float | Gaussian sigma for text detection | 1.5 |
angle |
float | Face rotation angle | 0 |
mask |
string | File path to the mask image used to retain specific image areas | |
rmask |
string | File path to the mask image used to force removing specific image areas. | |
color |
string | Seam color | #ff0000 |
shape |
string | Shape type used for debugging: circle,line |
circle |
A GUI interface is also incorporated into the library to visualize the process in real time. The GUI is based on the Gio library. Prior running it please make sure that you have installed all the required dependencies noted in the installation section (https://gioui.org/#installation) .
The preview window is activated by default but you can deactivate it any time by setting the -preview flag to false. When the images are processed concurrently from a directory the preview mode is deactivated by default.
The library is capable to detect human faces prior resizing the images by using the lightweight Pigo (https://github.com/esimov/pigo) face detection library.
The image below illustrates the application capabilities for human face detection prior resizing. It's clearly visible that with face detection activated the algorithm will avoid cropping pixels inside the detected faces, retaining the face zone unaltered.
| Original image | With face detection | Without face detection |
|---|---|---|
![]() |
![]() |
![]() |
In order to detect faces prior rescaling, use the -faceDetect flag. There is no need to provide a face classification file, since it's already embedded into the generated binary file. The sample code below will resize the provided image with 20%, but checks for human faces in order tot avoid face deformations.
For face detection related settings please check the Pigo documentation.
$ caire -in input.jpg -out output.jpg -faceDetect=1 -perc=1 -width=20Starting with v1.6.0 the library support text detection capabilities also. This is hidden behind a feature flag named -textDetect.
Once activated the text detector will scan the image to detect text areas with identical characteristics, and mark them as grouped components belonging to the same zone. Then the color intensity is increased for each detected text zone, which tricks the Sobel detector to treat them as important areas which shouldn't be altered by the seam carver.
Note this can generate false positive results also, so adapt the flags for each image specifically.
As a rule of thumb, try to avoid using the -textDetect flag when the image does not contains texts. Otherwise play with -gaussian, -cannyHigh and -cannyLow parameters. These will tell the Canny detector how to threat the gradient changes in edge detection.
Also keep in mind that text detection will drop the performance. You can use the -textDetInterval flag to limit how often the detector should be activated during the seam removal operation.
| With text detection | Without text detection |
|---|---|
![]() |
![]() |
Below is a screen recording of the text detector in action.
Resize-with-activated-text-detection.mp4
$ caire -in input.jpg -out output.jpgYou can also use stdin and stdout with -:
$ cat input/source.jpg | caire -in - -out - >out.jpgin and out default to - so you can also use:
$ cat input/source.jpg | caire >out.jpg
$ caire -out out.jpg < input/source.jpgYou can provide also an image URL for the -in flag or even use curl or wget as a pipe command in which case there is no need to use the -in flag.
$ caire -in <image_url> -out <output-folder>
$ curl -s <image_url> | caire > out.jpgThe library can also process multiple images from a directory concurrently. You have to provide only the source and the destination folder and the new width or height in this case.
$ caire -in <input_folder> -out <output-folder>There is no need to define the output file type, just use the correct extension and the library will encode the image to that specific type.
In case you wish to scale down the image by a specific percentage, it can be used the -perc boolean flag. In this case the values provided for the width and height are expressed in percentage and not pixel values. For example to reduce the image dimension by 20% both horizontally and vertically you can use the following command:
$ caire -in input/source.jpg -out ./out.jpg -perc=1 -width=20 -height=20 -debug=falseAlso the library supports the -square option. When this option is used the image will be resized to a square, based on the shortest edge.
When an image is resized on both the X and Y axis, the algorithm will first try to rescale it prior resizing, but also will preserve the image aspect ratio. The seam carving algorithm is applied only to the remaining points. Ex. : given an image of dimensions 2048x1536 if we want to resize to the 1024x500, the tool first rescale the image to 1024x768 and then will remove only the remaining 268px.
-mask: The path to the protective mask. The mask should be in binary format and have the same size as the input image. White areas represent regions where no seams should be carved.-rmask: The path to the removal mask. The mask should be in binary format and have the same size as the input image. White areas represent regions to be removed.
| Mask protection | Mask removal |
|---|---|
mask.mp4 |
rmask.mp4 |
- Caire can be used as a serverless function via OpenFaaS: https://github.com/esimov/caire-openfaas
- Caire can also be used as a
snapfunction (https://snapcraft.io/caire):$ snap run caire --h
| Original | Shrunk |
|---|---|
![]() |
![]() |
![]() |
![]() |
![]() |
![]() |
![]() |
![]() |
| Original | Extended |
|---|---|
![]() |
![]() |
![]() |
![]() |
- https://en.wikipedia.org/wiki/Seam_carving
- https://inst.eecs.berkeley.edu/~cs194-26/fa16/hw/proj4-seamcarving/imret.pdf
- http://pages.cs.wisc.edu/~moayad/cs766/download_files/alnammi_cs_766_final_report.pdf
- https://stacks.stanford.edu/file/druid:my512gb2187/Zargham_Nassirpour_Content_aware_image_resizing.pdf
- Endre Simo (@simo_endre)
Copyright © 2018 Endre Simo
This project is under the MIT License. See the LICENSE file for the full license text.





















