Image segmentation web service designed for hosting on Heroku.
The API is inspired by imgix - you pass in the image url (optionally urlencoded), control the algorithm with query parameters, and recieve a segmented image as the response (or a JSON formatted error). This allows you to embed it in a web page, or play with the algorithm in your browser without having to install any software.
https://segment-image.herokuapp.com/<image_url>?cluster_method=kmeans&num_clusters=10The image above on the right is generated by the following URL (open it in your browser and try changing the parameters!):
https://segment-image.herokuapp.com/http://i.imgur.com/hsiR4hV.jpg?cluster_method=meanshift&quantile=0.015&colour_space=ycrcb
The service supports the following parameters:
| Parameter | Description |
|---|---|
cluster_method |
ward, meanshift or kmeans |
colour_space |
rgb, hsv, hls, ycrcb, lab or luv |
num_clusters |
integer between 1 and 100, e.g. 5 (required with ward and kmeans) |
quantile |
a float between 0 and 1, e.g. 0.01 (required with meanshift) |
Ward is interesting, because it's configured with connectivity constraints - it'll cluster colours together by region. It's also the slowest algorithm, so may timeout (after ~25 seconds).
Configure Redis to evict the less recently used keys first:
heroku redis:maxmemory --policy allkeys-lru
| Method | Meanshift quantile=0.07 |
Meanshift quantile=0.15 |
|---|---|---|
| RGB |  |
 |
| HSV |  |
 |
| HLS |  |
 |
| YCrCb |  |
 |
| LAB |  |
 |
| LUV |  |
 |
| Method | K-Means num_clusters=4 |
K-means num_clusters=8 |
|---|---|---|
| RGB |  |
 |
| HSV |  |
 |
| HLS |  |
 |
| YCrCb |  |
 |
| LAB |  |
 |
| LUV |  |
 |
| Method | Ward num_clusters=6 |
Ward num_clusters=15 |
|---|---|---|
| RGB |  |
 |
| HSV |  |
 |
| HLS |  |
 |
| YCrCb |  |
 |
| LAB |  |
 |
| LUV |  |
 |