Problem
- What is the difference between resizeBilinear(*1) and resizeNearestNeighbor(*2)? In particular, resizeNearestNeighbor does not return the correct image (half the image is black). (Half of the image is black.)
- ResizeBilinear does not return the correct results for segmentation. (Half of the image is gray.) Why is that? This is very similar to the results of resizeNearestNeighbor.
Background
I want to develop an application that does segmentation in Tensorflow.js. Fortunately, I have found some sample code for segmentation in Python. (*3) I believe I can do segmentation in Tensorflow.js by converting the Keras model acquired there to the TFJS model. I have tried to resize the images in Tensorflow.js but I can't get the right results. Does anyone have any good ideas?
Code
Here's the source code I wrote. (I'm not used to writing JavaScript.)
<!DOCTYPE html>
<html>
<head></head>
<meta charset="utf-8"/>
<body>
<script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs"></script>
<canvas id="canvas1" width="256" height="256" style="border: 2px solid;"></canvas>
<canvas id='canvas2' width="128" height="128" style="border: 2px solid;"></canvas>
<canvas id='canvas3' width="128" height="128" style="border: 2px solid;"></canvas>
<canvas id='canvas4' width="128" height="128" style="border: 2px solid;"></canvas>
<canvas id='canvas5' width="128" height="128" style="border: 2px solid;"></canvas>
<script>
var inputImg = new Image();inputImg.src = "../00_resources/woman172.jpg";
const canvas1 = document.getElementById('canvas1');
const canvas2 = document.getElementById('canvas2');
const canvas3 = document.getElementById('canvas3');
const canvas4 = document.getElementById('canvas4');
const canvas5 = document.getElementById('canvas5');
const SIZE = 128;
const COL_CHANNEL = 3;
const BIN_CHANNEL = 1;
inputImg.onload = () => {
const inCtx = canvas1.getContext('2d');
inCtx.drawImage(inputImg, 0, 0, canvas1.width, canvas1.height);
};
(async() => {
// load model
let model = await tf.loadGraphModel('/tfjsdoc/Portrait-Segmentation/model/deconv_bnoptimized_munet_to_tfjs_graph_model/model.json');
// image1 (original image)
let img1 = tf.browser.fromPixels(canvas1, COL_CHANNEL);
tf.browser.toPixels(img1, canvas2);// OK
// image2 (resized image - using 'resizeNearestNeighbor')
let img2 = img1.resizeNearestNeighbor([SIZE, SIZE]).toFloat().div(tf.scalar(255));
tf.browser.toPixels(img2, canvas3);// NG
// image3 (resized image - using 'resizeBilinear')
let img3 = img1.resizeBilinear([SIZE, SIZE]).toFloat().div(tf.scalar(255));
tf.browser.toPixels(img3, canvas4);// OK?
// predict (image3 is used to make predictions, but the same problem occurs as in image2)
let pred = await model.predict(tf.reshape(img3, [-1,SIZE,SIZE,COL_CHANNEL])).data();
tf.browser.toPixels(tf.tensor(pred, [SIZE,SIZE, BIN_CHANNEL]), canvas5);// NG
})();
</script>
</body>
</html>
Capturing the Execution Result
From left to right.
- original image
- The original image is taken and drawn on canvas
- Original image resized using resizeNearestNeighbor and drawn on canvas (half of the image is black)
- Original image resized using resizeBilinear and drawn on canvas (looks right)
- Segmentation of the original image resized using resizeBilinear and drawn on canvas (half of the image is gray)
References
(*1) https://js.tensorflow.org/api/latest/#image.resizeBilinear
(*2) https://js.tensorflow.org/api/latest/#image.resizeNearestNeighbor
From my experience, it seems like both accomplish similar resizing of the input image. ResizeBilinear performs a bilinear interpolation on the image.