javascript - How can I improve this WebGL / GLSL image downsampling shader

Question

I am using WebGL to resize images clientside very quickly within an app I am working on. I have written a GLSL shader that performs simple bilinear filtering on the images that I am downsizing.

It works fine for the most part but there are many occasions where the resize is huge e.g. from a 2048x2048 image down to 110x110 in order to generate a thumbnail. In these instances the quality is poor and far too blurry.

My current GLSL shader is as follows:

uniform float textureSizeWidth;
uniform float textureSizeHeight;
uniform float texelSizeX;
uniform float texelSizeY;
varying mediump vec2 texCoord;
uniform sampler2D texture;

vec4 tex2DBiLinear( sampler2D textureSampler_i, vec2 texCoord_i )
{
    vec4 p0q0 = texture2D(textureSampler_i, texCoord_i);
    vec4 p1q0 = texture2D(textureSampler_i, texCoord_i + vec2(texelSizeX, 0));

    vec4 p0q1 = texture2D(textureSampler_i, texCoord_i + vec2(0, texelSizeY));
    vec4 p1q1 = texture2D(textureSampler_i, texCoord_i + vec2(texelSizeX , texelSizeY));

    float a = fract( texCoord_i.x * textureSizeWidth );

    vec4 pInterp_q0 = mix( p0q0, p1q0, a );
    vec4 pInterp_q1 = mix( p0q1, p1q1, a );

    float b = fract( texCoord_i.y * textureSizeHeight );
    return mix( pInterp_q0, pInterp_q1, b );
}
void main() { 

    gl_FragColor = tex2DBiLinear(texture,texCoord);
}');

TexelsizeX and TexelsizeY are simply (1.0 / texture width) and height respectively...

I would like to implement a higher quality filtering technique, ideally a [Lancosz][1] filter which should produce far better results but I cannot seem to get my head around how to implement the algorithm with GLSL as I am very new to WebGL and GLSL in general.

Could anybody point me in the right direction?

Thanks in advance.

Answer 1

If you're looking for Lanczos resampling, the following is the shader program I use in my open source GPUImage library:

Vertex shader:

 attribute vec4 position;
 attribute vec2 inputTextureCoordinate;

 uniform float texelWidthOffset;
 uniform float texelHeightOffset;

 varying vec2 centerTextureCoordinate;
 varying vec2 oneStepLeftTextureCoordinate;
 varying vec2 twoStepsLeftTextureCoordinate;
 varying vec2 threeStepsLeftTextureCoordinate;
 varying vec2 fourStepsLeftTextureCoordinate;
 varying vec2 oneStepRightTextureCoordinate;
 varying vec2 twoStepsRightTextureCoordinate;
 varying vec2 threeStepsRightTextureCoordinate;
 varying vec2 fourStepsRightTextureCoordinate;

 void main()
 {
     gl_Position = position;

     vec2 firstOffset = vec2(texelWidthOffset, texelHeightOffset);
     vec2 secondOffset = vec2(2.0 * texelWidthOffset, 2.0 * texelHeightOffset);
     vec2 thirdOffset = vec2(3.0 * texelWidthOffset, 3.0 * texelHeightOffset);
     vec2 fourthOffset = vec2(4.0 * texelWidthOffset, 4.0 * texelHeightOffset);

     centerTextureCoordinate = inputTextureCoordinate;
     oneStepLeftTextureCoordinate = inputTextureCoordinate - firstOffset;
     twoStepsLeftTextureCoordinate = inputTextureCoordinate - secondOffset;
     threeStepsLeftTextureCoordinate = inputTextureCoordinate - thirdOffset;
     fourStepsLeftTextureCoordinate = inputTextureCoordinate - fourthOffset;
     oneStepRightTextureCoordinate = inputTextureCoordinate + firstOffset;
     twoStepsRightTextureCoordinate = inputTextureCoordinate + secondOffset;
     threeStepsRightTextureCoordinate = inputTextureCoordinate + thirdOffset;
     fourStepsRightTextureCoordinate = inputTextureCoordinate + fourthOffset;
 }

Fragment shader:

 precision highp float;

 uniform sampler2D inputImageTexture;

 varying vec2 centerTextureCoordinate;
 varying vec2 oneStepLeftTextureCoordinate;
 varying vec2 twoStepsLeftTextureCoordinate;
 varying vec2 threeStepsLeftTextureCoordinate;
 varying vec2 fourStepsLeftTextureCoordinate;
 varying vec2 oneStepRightTextureCoordinate;
 varying vec2 twoStepsRightTextureCoordinate;
 varying vec2 threeStepsRightTextureCoordinate;
 varying vec2 fourStepsRightTextureCoordinate;

 // sinc(x) * sinc(x/a) = (a * sin(pi * x) * sin(pi * x / a)) / (pi^2 * x^2)
 // Assuming a Lanczos constant of 2.0, and scaling values to max out at x = +/- 1.5

 void main()
 {
     lowp vec4 fragmentColor = texture2D(inputImageTexture, centerTextureCoordinate) * 0.38026;

     fragmentColor += texture2D(inputImageTexture, oneStepLeftTextureCoordinate) * 0.27667;
     fragmentColor += texture2D(inputImageTexture, oneStepRightTextureCoordinate) * 0.27667;

     fragmentColor += texture2D(inputImageTexture, twoStepsLeftTextureCoordinate) * 0.08074;
     fragmentColor += texture2D(inputImageTexture, twoStepsRightTextureCoordinate) * 0.08074;

     fragmentColor += texture2D(inputImageTexture, threeStepsLeftTextureCoordinate) * -0.02612;
     fragmentColor += texture2D(inputImageTexture, threeStepsRightTextureCoordinate) * -0.02612;

     fragmentColor += texture2D(inputImageTexture, fourStepsLeftTextureCoordinate) * -0.02143;
     fragmentColor += texture2D(inputImageTexture, fourStepsRightTextureCoordinate) * -0.02143;

     gl_FragColor = fragmentColor;
 }

This is applied in two passes, with the first performing a horizontal downsampling and the second a vertical downsampling. The texelWidthOffset and texelHeightOffset uniforms are alternately set to 0.0 and the width fraction or height fraction of a single pixel in the image.

I hard-calculate the texel offsets in the vertex shader because this avoids dependent texture reads on the mobile devices I'm targeting with this, leading to significantly better performance there. It is a little verbose, though.

Results from this Lanczos resampling:

Normal bilinear downsampling:

Nearest-neighbor downsampling: