BufferedImage's .getRaster .getNumBands can be used in figuring out the shape

[daslu]

The namespace tech.v3.libs.buffered-image extends BufferedImage with certain protocols, including the extraction of shape (tech.v3.datatype/shape):
https://github.com/cnuernber/dtype-next/blob/9170e69/src/tech/v3/libs/buffered_image.clj#L160

In unknown image types, it is assumed only one band:
https://github.com/cnuernber/dtype-next/blob/9170e69/src/tech/v3/libs/buffered_image.clj#L177

However, the number of bands can be extracted using .getRaster .getNumBands, as @larzeitlin does in an upcoming tutorial about Cloud Optimized GeoTiff:
https://github.com/ClojureCivitas/clojurecivitas.github.io/blob/be26c88/src/gis/geotiff.clj#L135

Example test file: https://github.com/ClojureCivitas/clojurecivitas.github.io/blob/be26c887796258f9ae31756cc3da7b654cf8d11c/src/gis/resources/example_geotiff_medium.tif

You can download it, and use the following code from @larzeitlin's tutorial to read it and extract the number of bands:

(import '[java.awt.image BufferedImage]
        '[java.io File]
        '[javax.imageio ImageIO]
        '[javax.imageio.stream FileImageInputStream])

(defn tiff-reader []
  (let [readers (ImageIO/getImageReadersByFormatName "tiff")]
    (.next readers)))

(defn read-tiff [file-path]
  (let [file (File. file-path)
        reader (tiff-reader)]
    (with-open [stream (FileImageInputStream. file)]
      (.setInput reader stream)
      (let [images (doall
                    ;; A single TIFF can store multiple images.
                    ;; This will be significant when we come to
                    ;; GeoTIFFs because of resolution pyramids.
                    ;; See below.
                    (for [i (range (.getNumImages reader true))]
                      (let [metadata (.getImageMetadata reader i)
                            native-format (.getNativeMetadataFormatName metadata)]
                        {:index i
                         :native-format native-format
                         :width (.getWidth reader i)
                         :height (.getHeight reader i)
                         :image (.read reader i)
                         :metadata metadata})))]
        {:images images
         :reader-format (.getFormatName reader)}))))

(-> "example_geotiff_medium.tif"
    read-tiff
    :images
    first
    :image
    .getRaster
    .getNumBands)

This will also affect tech.v3.tensor/as-tensor and allow it to come up with the right shape of output, this allowing to turn any Raster image to a tensor -- quite important for working with multispectral images.

Zulip discussion with @harold: #tech.ml.dataset.dev > shape of an image

[harold]

Thanks @daslu - this is great.

The example_geotiff_medium.tif seems to be sort of big (80+MB), and not multi-image or multispectral.

Though it does load and shape returns [4999 5999 1]

I am working on a patch that makes shape return [4999 5999 3], which is an improvement, but I'd like to test it against additional files.

Do you know of any other test images? Perhaps smaller, and maybe w/ more than one image per file, or more than 3 bands per image?

[larzeitlin]

example geotiff: https://github.com/larzeitlin/example_small_ms_geotiff

[harold]

@larzeitlin - super 🆒, well done, we'll check it out.

[benkamphaus]

AFAIK this should support multiplex ome-tiff, too (open microscopy format). There's example data here. getRaster -> getNumBands appears to be used heavily in QuPath w/o much exception babying. [ ex ]

[harold]

AFAIK this should support multiplex ome-tiff, too (open microscopy format). There's example data here. getRaster -> getNumBands appears to be used heavily in QuPath w/o much exception babying. [ ex ]

Great to see you here, the whole buffered-image namespace has roots in our original conversations about multispectral images, so it's nice to see this progressing. This is on our list to look at, thanks for everyone's patience.