VipsImageconv

Perform a convolution of in with mask.

Each output pixel is calculated as:

sigma[i]{pixel[i] * mask[i]} / scale + offset

where scale and offset are part of mask.

By default, precision is VIPS_PRECISION_FLOAT. The output image is always VIPS_FORMAT_FLOAT unless in is VIPS_FORMAT_DOUBLE, in which case out is also VIPS_FORMAT_DOUBLE.

If precision is VIPS_PRECISION_INTEGER, then elements of mask are converted to integers before convolution, using rint(), and the output image always has the same VipsBandFormat as the input image.

For VIPS_FORMAT_UCHAR images and VIPS_PRECISION_INTEGER precision, vips_conv() uses a fast vector path based on half-float arithmetic. This can produce slightly different results. Disable the vector path with --vips-novector or VIPS_NOVECTOR or vips_vector_set_enabled().

If precision is VIPS_PRECISION_APPROXIMATE then, like VIPS_PRECISION_INTEGER, mask is converted to int before convolution, and the output image always has the same VipsBandFormat as the input image.

Larger values for layers give more accurate results, but are slower. As layers approaches the mask radius, the accuracy will become close to exact convolution and the speed will drop to match. For many large masks, such as Gaussian, n_layers need be only 10% of this value and accuracy will still be good.

Smaller values of cluster will give more accurate results, but be slower and use more memory. 10% of the mask radius is a good rule of thumb.

This method is not directly available to language bindings.