Andy Grundman > Image-Scale-0.06 > Image::Scale



Annotate this POD


Open  2
View/Report Bugs
Module Version: 0.06   Source   Latest Release: Image-Scale-0.11


Image::Scale - Fast, high-quality fixed-point image resizing


    use Image::Scale
    # Resize to 150 width and save to a file
    my $img = Image::Scale->new('image.jpg') || die "Invalid JPEG file";
    $img->resize_gd( { width => 150 } );
    # Easily resize artwork embedded within an audio file
    # You can use L<Audio::Scan> to obtain offset/length information
    my $img = Image::Scale->new( 'track.mp3', { offset => 2200, length => 34123 } );
    $img->resize_gd_fixed_point( { width => 75, height => 75, keep_aspect => 1 } );
    my $data = $img->as_png();


This module implements several resizing algorithms with a focus on low overhead, speed and minimal features. Algorithms available are:

  GD's copyResampled (floating-point)
  GD's copyResampled fixed-point (useful on embedded devices/NAS devices)
  GraphicsMagick's assortment of resize filters (floating-point)
  GraphicsMagick's Triangle filter in fixed-point

Supported image formats include JPEG, GIF, PNG, and BMP for input, and JPEG and PNG for output.

This module came about because we needed to improve the very slow performance of floating-point resizing algorithms on platforms without a floating-point unit, such as ARM devices like the SheevaPlug, and the Sparc-based ReadyNAS Duo. Previously it would take many seconds to resize using GD on the ReadyNAS but the conversion to fixed-point with a little assembly code brings this down to the range of well under 1 second.

GD is also incredibly difficult to build on platforms such as Windows so we needed a replacement.

Normal platforms will also see improvement, by removing all of the GD overhead this version of copyResampled is around 3 times faster while also using less memory.

The fixed-point versions have an accuracy to around 4 decimal places so the quality of floating-point vs. fixed is essentially identical.


new( $PATH or \$DATA, [ \%OPTIONS ] )

Initialize a new Image::Scale object from PATH, which may be any valid JPEG, GIF, PNG, or BMP file.

Raw image data may also be passed in as a scalar reference. Using a file path is recommended when possible as this is more efficient and requires less memory.

new() reads the image header, and will return undef if the header is invalid, so be sure to check for this.

Optionally you can also pass in additional options in a hashref:


To access an image embedded within another file, such as an audio file, you can specify a byte offset and length.


Returns the width of the original source image.


Returns the height of the original source image.

resize( \%OPTIONS )

resize() uses the default resize algorithm, which is resize_gd_fixed_point. See below for details on the available options.

resize_gd( \%OPTIONS )

resize_gd_fixed_point( \%OPTIONS )

resize_gm( \%OPTIONS )

resize_gm_fixed_point( \%OPTIONS )

The 4 resize methods available are:

    resize_gd - This is GD's copyResampled algorithm (floating-point)
    resize_gd_fixed_point - copyResampled (converted to fixed-point)
    resize_gm - GraphicsMagick, see below for filter options
    resize_gm_fixed_point - GraphicsMagick, only the Triangle filter is available in fixed-point mode

Options are specified in a hashref:


At least one of width or height are required. If only one is supplied the image will retain the original aspect ratio.


For use with resize_gm() only. Choose from the following filters, sorted in order from least to most CPU time. This does not necessarily mean least to best quality, though! Be sure to do your own comparisons for quality.


If no filter is specified the default is Lanczos if downsizing, and Mitchell for upsizing or if the image has an alpha channel.

    keep_aspect => 1

Only useful when both width and height are specified. This option will keep the original aspect ratio of the source as well as center the image when resizing into a different aspect ratio. For best results, images altered in this way should be saved as PNG which will automatically add the necessary transparency around the image.

    bgcolor => 0xffffff

When using keep_aspect, you can use bgcolor to define the background color of the padded portion of the image. Usually this should only be used if saving as JPEG because PNG will default to transparent. If this value is set and the image is saved as PNG, the PNG will not be transparent. The default bgcolor value is 0x000000 (black).

    ignore_exif => 1

By default, if a JPEG image contains an EXIF tag with orientation info, the image will be rotated accordingly during resizing. To disable this feature, set ignore_exif to 1.

    memory_limit => $limit_in_bytes

To avoid excess memory growth when resizing images that may be very large, you can specify this option. If the resize_*() method would result in a total memory allocation greater than $limit_in_bytes, the method will die. Be sure to wrap the resize call in an eval when using this option.

save_jpeg( $PATH, [ $QUALITY ] )

Saves the resized image as a JPEG to PATH. If a quality is not specified, the quality defaults to 90.

as_jpeg( [ $QUALITY ] )

Returns the resized JPEG image as scalar data. If a quality is not specified, the quality defaults to 90.

save_png( $PATH )

Saves the resized image as a PNG to PATH. Transparency is preserved when saving to PNG.


Returns the resized PNG image as scalar data.




Returns the version of the image library used. Returns undef if support for that image format was not built.


These numbers were gathered on my 2.4ghz MacBook Pro.

JPEG image, 1425x1425 -> 200x200 (libjpeg v8 with scaling)

    GD copyResampled                        21.9/s
    resize_gm( { filter => 'Triangle' } )   65.7/s
    resize_gd_fixed_point                   67.9/s
    resize_gd                               69.4/s
    resize_gm_fixed_point                   74.5/s

PNG image, 512x768 -> 200x133 (libpng 1.4.3)

    GD copyResampled                        14.7/s
    resize_gm( { filter => 'Triangle' } )   26.2/s
    resize_gm_fixed_point                   27.7/s
    resize_gd                               29.9/s
    resize_gd_fixed_point                   31.9/s

Here are some numbers from a machine without floating-point support. (Marvell SheevaPlug 1.2ghz ARM9, JPEG 1425x1425 -> 200x200, libjpeg 6b with scaling)

    GD copyResampled                        1.08/s
    resize_gd                               2.16/s
    resize_gm( { filter => 'Triangle' } )   2.85/s
    resize_gd_fixed_point                   7.98/s
    resize_gm_fixed_point                   9.44/s

And finally, from an even slower machine, the 240mhz Netgear ReadyNAS Duo which has extremely poor floating-point performance. (JPEG 1425x1425 -> 200x200, libjpeg 6b with scaling)

    resize_gd                               0.029/s (34.5 s/iter)
    resize_gm( { filter => 'Triangle' } )   0.033/s (30.4 s/iter)
    resize_gd_fixed_point                   1.92/s  (0.522 s/iter)
    resize_gm_fixed_point                   2.07/s  (0.483 s/iter) (63x faster than floating-point!)


GD, Image::Magick, Imager


Andy Grundman, <>


Copyright (C) 2010 Andy Grundman

This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.

syntax highlighting: