In the early ML Kit example there was an example that used 'org.deeplearning4j' to easily convert an image to an in-memory tflite model:
tfLiteModel model = tf.keras.convert(JPEGimage);
sure you can train Machine Learning with single images, but you have to retrain once images change.
in above sample you could just replace the whole image.
i don't find the example anymore, it has not been moved to github.
is there any solution similarily easy?
Accepted Answer
That example probably never existed in that form. org.deeplearning4j is not a TFLite converter, and tf.keras.convert(image) is not a real call. A single image can't become a trained model anyway.
What you likely want is matching against a reference image you can swap without retraining. Use image embeddings: a pretrained model turns each image into a vector, MediaPipe's Image Embedder does this:
val options = ImageEmbedder.ImageEmbedderOptions.builder()
.setBaseOptions(BaseOptions.builder().setModelAssetPath("mobilenet_v3_small.tflite").build())
.setL2Normalize(true)
.build()
val embedder = ImageEmbedder.createFromOptions(context, options)
fun embed(bmp: Bitmap) =
embedder.embed(BitmapImageBuilder(bmp).build()).embeddingResult().embeddings().first()
var reference = embed(referenceBitmap) // swap anytime, no retraining
fun matches(bmp: Bitmap) =
ImageEmbedder.cosineSimilarity(reference, embed(bmp)) > 0.8 // tune the threshold
If you need an exact match rather than a similar-looking one, OpenCV feature matching (ORB/SIFT) may work better. I haven’t run this, so test the threshold on your own images.
Tanveer Munir