Heavy image downsampling artefacts

问题

I need to heavily downsample a image to prepare for a perceptual hashing algorithm, for example from a 3024x4032 to a 9x9 image, and I'd like to use webgl for that (as the base image is already processed within the webgl pipeline), and artefacts quickly appear when the downsampling rate is too large.

Taking this sample situation:

// WebGL2 - 2D Image
// from https://webgl2fundamentals.org/webgl/webgl-2d-image.html

"use strict";

var downscaleFactor = 0.5;
var vertexShaderSource = `#version 300 es

// an attribute is an input (in) to a vertex shader.
// It will receive data from a buffer
in vec2 a_position;
in vec2 a_texCoord;

// Used to pass in the resolution of the canvas
uniform vec2 u_resolution;

// Used to pass the texture coordinates to the fragment shader
out vec2 v_texCoord;

// all shaders have a main function
void main() {

  // convert the position from pixels to 0.0 to 1.0
  vec2 zeroToOne = a_position / u_resolution;

  // convert from 0->1 to 0->2
  vec2 zeroToTwo = zeroToOne * 2.0;

  // convert from 0->2 to -1->+1 (clipspace)
  vec2 clipSpace = zeroToTwo - 1.0;

  gl_Position = vec4(clipSpace * vec2(1, -1), 0, 1);

  // pass the texCoord to the fragment shader
  // The GPU will interpolate this value between points.
  v_texCoord = a_texCoord;
}
`;

var fragmentShaderSource = `#version 300 es

// fragment shaders don't have a default precision so we need
// to pick one. mediump is a good default. It means "medium precision"
precision mediump float;

// our texture
uniform sampler2D u_image;

// the texCoords passed in from the vertex shader.
in vec2 v_texCoord;

// we need to declare an output for the fragment shader
out vec4 outColor;

void main() {
  outColor = texture(u_image, v_texCoord);
}
`;

var image = new Image();
image.src = "https://i.imgur.com/AetKHN1.png";
image.crossOrigin = "anonymous";
image.onload = function() {
  render(image);
};

function render(image) {
  // Get A WebGL context
  /** @type {HTMLCanvasElement} */
  var canvas = document.getElementById("canvas");
  var gl = canvas.getContext("webgl2");
  if (!gl) {
    return;
  }

  // setup GLSL program
  var program = webglUtils.createProgramFromSources(gl, [vertexShaderSource, fragmentShaderSource]);

  // look up where the vertex data needs to go.
  var positionAttributeLocation = gl.getAttribLocation(program, "a_position");
  var texCoordAttributeLocation = gl.getAttribLocation(program, "a_texCoord");

  // lookup uniforms
  var resolutionLocation = gl.getUniformLocation(program, "u_resolution");
  var imageLocation = gl.getUniformLocation(program, "u_image");

  // Create a vertex array object (attribute state)
  var vao = gl.createVertexArray();

  // and make it the one we're currently working with
  gl.bindVertexArray(vao);

  // Create a buffer and put a single pixel space rectangle in
  // it (2 triangles)
  var positionBuffer = gl.createBuffer();

  // Turn on the attribute
  gl.enableVertexAttribArray(positionAttributeLocation);

  // Bind it to ARRAY_BUFFER (think of it as ARRAY_BUFFER = positionBuffer)
  gl.bindBuffer(gl.ARRAY_BUFFER, positionBuffer);

  // Tell the attribute how to get data out of positionBuffer (ARRAY_BUFFER)
  var size = 2; // 2 components per iteration
  var type = gl.FLOAT; // the data is 32bit floats
  var normalize = false; // don't normalize the data
  var stride = 0; // 0 = move forward size * sizeof(type) each iteration to get the next position
  var offset = 0; // start at the beginning of the buffer
  gl.vertexAttribPointer(
    positionAttributeLocation, size, type, normalize, stride, offset);

  // provide texture coordinates for the rectangle.
  var texCoordBuffer = gl.createBuffer();
  gl.bindBuffer(gl.ARRAY_BUFFER, texCoordBuffer);
  gl.bufferData(gl.ARRAY_BUFFER, new Float32Array([
    0.0, 0.0,
    1.0, 0.0,
    0.0, 1.0,
    0.0, 1.0,
    1.0, 0.0,
    1.0, 1.0,
  ]), gl.STATIC_DRAW);
  gl.enableVertexAttribArray(texCoordAttributeLocation);
  var size = 2; // 2 components per iteration
  var type = gl.FLOAT; // the data is 32bit floats
  var normalize = false; // don't normalize the data
  var stride = 0; // 0 = move forward size * sizeof(type) each iteration to get the next position
  var offset = 0; // start at the beginning of the buffer
  gl.vertexAttribPointer(
    texCoordAttributeLocation, size, type, normalize, stride, offset);

  // Create a texture.
  var texture = gl.createTexture();

  // make unit 0 the active texture uint
  // (ie, the unit all other texture commands will affect
  gl.activeTexture(gl.TEXTURE0 + 0);

  // Bind it to texture unit 0' 2D bind point
  gl.bindTexture(gl.TEXTURE_2D, texture);

  // Set the parameters so we don't need mips and so we're not filtering
  // and we don't repeat at the edges
  gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_WRAP_S, gl.CLAMP_TO_EDGE);
  gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_WRAP_T, gl.CLAMP_TO_EDGE);
  gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_MIN_FILTER, gl.LINEAR);
  gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_MAG_FILTER, gl.LINEAR);

  // Upload the image into the texture.
  var mipLevel = 0; // the largest mip
  var internalFormat = gl.RGBA; // format we want in the texture
  var srcFormat = gl.RGBA; // format of data we are supplying
  var srcType = gl.UNSIGNED_BYTE; // type of data we are supplying
  gl.texImage2D(gl.TEXTURE_2D,
    mipLevel,
    internalFormat,
    srcFormat,
    srcType,
    image);

  webglUtils.resizeCanvasToDisplaySize(gl.canvas);

  // Tell WebGL how to convert from clip space to pixels
  gl.viewport(0, 0, gl.canvas.width, gl.canvas.height);

  // Clear the canvas
  gl.clearColor(0, 0, 0, 0);
  gl.clear(gl.COLOR_BUFFER_BIT | gl.DEPTH_BUFFER_BIT);

  // Tell it to use our program (pair of shaders)
  gl.useProgram(program);

  // Bind the attribute/buffer set we want.
  gl.bindVertexArray(vao);

  // Pass in the canvas resolution so we can convert from
  // pixels to clipspace in the shader
  gl.uniform2f(resolutionLocation, gl.canvas.width, gl.canvas.height);

  // Tell the shader to get the texture from texture unit 0
  gl.uniform1i(imageLocation, 0);

  // Bind the position buffer so gl.bufferData that will be called
  // in setRectangle puts data in the position buffer
  gl.bindBuffer(gl.ARRAY_BUFFER, positionBuffer);

  // Set a rectangle the same size as the image.
  setRectangle(gl, 0, 0, downscaleFactor * image.width, downscaleFactor * image.height);

  // Draw the rectangle.
  var primitiveType = gl.TRIANGLES;
  var offset = 0;
  var count = 6;
  gl.drawArrays(primitiveType, offset, count);
}

function setRectangle(gl, x, y, width, height) {
  var x1 = x;
  var x2 = x + width;
  var y1 = y;
  var y2 = y + height;
  gl.bufferData(gl.ARRAY_BUFFER, new Float32Array([
    x1, y1,
    x2, y1,
    x1, y2,
    x1, y2,
    x2, y1,
    x2, y2,
  ]), gl.STATIC_DRAW);
}

body {
  margin: 0;
}

canvas {
  width: 500px;
  height: 500px;
  display: block;
}

<canvas id="canvas"></canvas>

<!--
for most samples webgl-utils only provides shader compiling/linking and
canvas resizing because why clutter the examples with code that's the same in every sample.
See http://webgl2fundamentals.org/webgl/lessons/webgl-boilerplate.html
and http://webgl2fundamentals.org/webgl/lessons/webgl-resizing-the-canvas.html
for webgl-utils, m3, m4, and webgl-lessons-ui.
-->
<script src="https://webgl2fundamentals.org/webgl/resources/webgl-utils.js"></script>

where the base image is set-up with gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_MIN_FILTER, gl.LINEAR);, the image resize works fine with downscaleFactor >= 0.5, until I reduce the factor further when artefacts appear and things get ugly (try downscaleFactor = 0.15 at line 6 and run).

I've tried changing the fragment shader's texture with a bilinear interpolation function, but the issue still exists.

vec4 textureBilinear(sampler2D textureSampler, vec2 textureCoordinate) {
  vec2 texelSize = 1.0 / vec2(textureSize(textureSampler, 0));
  vec4 p0q0 = texture(textureSampler, textureCoordinate);
  vec4 p1q0 = texture(textureSampler, textureCoordinate + vec2(texelSize.x, 0));
  vec4 p0q1 = texture(textureSampler, textureCoordinate + vec2(0, texelSize.y));
  vec4 p1q1 = texture(textureSampler, textureCoordinate + vec2(texelSize.x , texelSize.y));

  float a = fract(textureCoordinate.x);
  float b = fract(textureCoordinate.y);
  vec4 pInterp_q0 = mix(p0q0, p1q0, a);
  vec4 pInterp_q1 = mix(p0q1, p1q1, a);
  return mix(pInterp_q0, pInterp_q1, b);
}

When I do try to downsample the image with a photo-editing software, I get a smooth result as expected (here, bilinear). What are strategies that would yield the same result? (possibly in one pass)

来源：https://stackoverflow.com/questions/54227348/heavy-image-downsampling-artefacts