Libnetpbm Image Processing Manual

Updated: December 2003

Table Of Contents

This reference manual covers functions in the libnetpbm library for processing images, using the Netpbm image formats and the libnetpbm in-memory image formats.

For historical reasons as well as to avoid clutter, it does not cover the largely obsolete PBM, PGM, PPM, and PNM classes of libnetpbm functions. For those, see PBM Function Manual, PGM Function Manual, PPM Function Manual, and PNM Function Manual. Note that you do not need those functions to process PBM, PGM, PPM, and PNM images. The functions in this manual are sufficient for that.

The PPM drawing functions are covered separately in PPM Drawing Function Manual.

For introductory and general information using libnetpbm, see Libnetpbm User's Guide.

libnetpbm also contains functions that are not specifically oriented toward processing image data. Read about those in the Libnetpbm Utility Manual.

To use these services, #include pam.h.


Here are some important types that you use with libnetpbm:

A sample of a Netpbm image. See the format specifications -- as an example, the red intensity of a particular pixel of a PPM image is a sample. This is an integer type.
A tuple from a PAM image or the PAM equivalent of a PNM image. See the PAM format specification -- as an example, a pixel of a PPM image would be a tuple. A tuple is an array of samples.
Same as sample, except in normalized form. This is a floating point type with a value in the range 0..1. 0 corresponds to a PAM/PNM sample value of 0. 1 corresponds to a PAM/PNM sample value equal to the image's maxval.
The same as tuple, except composed of normalized samples (samplen) instead of regular samples (sample).

struct pam

The main argument to most of the PAM functions is the address of a pam structure, which is defined as follows:

    struct pam {
        int size
        int len
        FILE *file   
        int format
        int plainformat
        int height
        int width
        int depth
        sample maxval
        int bytes_per_sample
        char tuple_type[256]
        int allocation_depth
        char **comment_p;

See The Libnetbm User's Guide for information on the pam structure.


PNM_MAXMAXVAL is the maximum maxval that Netpbm images could historically have: 255. Many programs aren't capable of handling Netpbm images with a maxval larger than this. It's named this way for backward compatibility -- it had this name back when it was the maximum maxval.

PNM_OVERALLMAXVAL is the maximum maxval that Netpbm images can have today (65535).

PBM_FORMAT, RPBM_FORMAT, PGM_FORMAT, RPGM_FORMAT, PPM_FORMAT, RPPM_FORMAT, and PAM_FORMAT are the format codes of the various Netpbm formats. RPBM_FORMAT is the raw PBM format and PBM_FORMAT is the plain PBM format, and so on. See the format member of the pam structure.

PAM_FORMAT_TYPE(format) gives the type of a format, given the format code. The types of formats are PBM, PGM, PPM, and PAM and macros for the type codes are, respectively, PBM_TYPE, PGM_TYPE, PPM_TYPE, and PAM_TYPE. Note that there are more format codes then there are format types because there are different format codes for the plain and raw subformats of each format.

Macros for the tuple types that are defined by Netpbm are as follows. See the tuple_type member of the pam structure.


These interfaces are declared in pam.h.

Memory Management


tuple ** pnm_allocpamarray( struct pam *pamP);

tuple * pnm_allocpamrow( struct pam *pamP);

void pnm_freepamarray( tuple **tuplearray, struct pam *pamP);

void pnm_freepamrow( tuple *tuplerow);

tuple * allocpamtuple( struct pam *pamP);

void pnm_freepamtuple( tuple tuple );

tuplen * pnm_allocpamrown( struct pam *pamP);

void pnm_freepamrown( tuplen *tuplenrow);


pnm_allocpamarray() allocates space for an array of tuples. pnm_freepamarray() frees an array space allocated by pnm_allocpamarray() or pnm_readpam().

pnm_allocpamrow() allocates space for a row of a PAM image, in basic form. pnm_freepamrow() frees it.

pnm_allocpamrown() is the same as pnm_allocpamrow() except that it allocates space for a PAM row in the normalized form. pnm_freepamrown() is similarly like pnm_freepamrow.

Reading Netpbm Files


void pnm_readpaminit( FILE *file, struct pam *pamP, int size);

void pnm_readpamrow( struct pam *pamP, tuple *tuplerow);

tuple ** pnm_readpam( FILE *file, struct pam *pamP, int size);

void pnm_readpamrown( struct pam *pamP, tuplen *tuplenrow);


pnm_readpaminit() reads the header of a Netpbm image.

See above for a general description of the pamP argument.

pnm_readpaminit() returns the information from the header in the *pamP structure. It does not require any members of *pamP through tuple_type to be set at invocation, and sets all of those members. It expects all members after tuple_type to be meaningful.

size is the size of the *pamP structure as understood by the program processing the image. pnm_readpaminit() does not attempt to use or set any members of the structure beyond that. The point of this argument is that the definition of the structure may change over time, with additional fields being added to the end. This argument allows pnm_readpaminit to distinguish between a new program that wants to exploit the additional features and an old program that cannot (or a new program that just doesn't want to deal with the added complexity). At a minimum, this size must contain the members up through tuple_type. You should use the PAM_STRUCT_SIZE macro to compute this argument. E.g. PAM_STRUCT_SIZE(tuple_type). PAM_STRUCT_SIZE was introduced in Netpbm 10.23 (July 2004). In older Netpbm, you can just use sizeof(), but then your code is not forward compatible at the source code level with newer libnetpbm (because when you compile it with newer libnetpbm header files, you'll be saying your structure contains all the new members that have been invented, but your code doesn't actually initialize them). So you might want to compute a proper size yourself.

The function expects to find the image file positioned to the start of the header and leaves it positioned to the start of the raster.

pnm_readpamrow() reads a row of the raster from a Netpbm image file. It expects all of the members of the *pamP structure to be set upon invocation and does not modify any of them. It expects to find the file positioned to the start of the row in question in the raster and leaves it positioned just after it. It returns the row as the array of tuples tuplerow, which must already have its column pointers set up so that it forms a C 2-dimensional array. The leftmost tuple is Element 0 of this array.

pnm_readpam() reads an entire image from a PAM or PNM image file and allocates the space in which to return the raster. It expects to find the file positioned to the first byte of the image and leaves it positioned just after the image.

*pamP is the same as for pnm_readpaminit().

The return value is a newly allocated array of the rows of the image, with the top row being Element 0 of the array. Each row is represented as pnm_readpamrow() would return.

The return value is also effectively a 3-dimensional C array of samples, with the dimensions corresponding to the height, width, and depth of the image, in that order.

pnm_readpam() combines the functions of pnm_allocpamarray(), pnm_readpaminit(), and iterations of pnm_readpamrow(). It may require more dynamic storage than you can afford.

pnm_readpamrown() is like pnm_readpamrow() except that it returns the row contents in normalized form (composed of normalized tuples (tuplen) instead of basic form (tuple).

pnm_readpaminit() and pnm_readpam abort the program with a message to Standard Error if the PAM or PNM image header is not syntactically valid, including if it contains a number too large to be processed using the system's normal data structures (to wit, a number that won't fit in a C 'int').

Writing Netpbm Files


void pnm_writepaminit( struct pam *pamP);

void pnm_writepamrow( struct pam *pamP, const tuple *tuplerow);

void pnm_writepam( struct pam *pamP, const tuple * const *tuplearray);

void pnm_writepamrown( struct pam *pamP, const tuplen *tuplerown);

void pnm_formatpamrow( struct pam *pamP, const tuple *tuplerow unsigned char * const outbuf, unsigned int * const rowSizeP );


pnm_writepaminit() writes the header of a PAM or PNM image and computes some of the fields of the pam structure.

See above for a description of the pamP argument.

The following members of the *pamP structure must be set upon invocation to tell the function how and what to write. size, len, file, format, height, width, depth, maxval. Furthermore, if format is PAM_FORMAT, tuple_type must be set and if format is not PAM_FORMAT, plainformat must be set.

pnm_writepaminit() sets the bytes_per_sample member based on the information supplied.

pnm_writepamrow() writes a row of the raster into a PAM or PNM image file. It expects to find the file positioned where the row should start and leaves it positioned just after the row. The function requires all the elements of *pamP to be set upon invocation and doesn't modify them.

tuplerow is an array of tuples representing the row. The leftmost tuple is Element 0 of this array.

pnm_writepam() writes an entire PAM or PNM image to a PAM or PNM image file. It expects to find the file positioned to where the image should start and leaves it positioned just after the image.

The members of the *pamP structure that must be set up invocation, and their meanings, is the same as for pnm_writepaminit.

pnm_writepam() sets the bytes_per_sample member based on the information supplied.

tuplearray is an array of rows such that you would pass to pnm_writepamrow(), with the top row being Element 0 of the array.

pnm_writepam() combines the functions of pnm_writepaminit(), and iterations of pnm_writepamrow(). Its raster input may be more storage than you can afford.

pnm_writepamrown() is like pnm_writepamrow() except that it takes the row contents in normalized form (composed of normalized tuples (tuplen) instead of basic form (tuple).

pnm_formatpamrow() is like pnm_writepamrow(), except that instead of writing a row to a file, it places the same bytes that would go in the file in a buffer you supply. There isn't an equivalent function to construct an image header; i.e. there is no analog to pnm_writepaminit(). But the header format, particularly for PAM, is so simple that you can easily build it yourself with standard C library string functions.

pnm_formatpamrow() was new in Netpbm 10.25 (October 2004).

Transforming Pixels


void pnm_YCbCrtuple( tuple tuple, double *YP, double *CrP, double *CbP);

void pnm_YCbCr_to_rgbtuple( const struct pam * const pamP, tuple const tuple, double const Y, double const Cb, double const Cr, int * const overflowP);

extern double pnm_lumin_factor[3];

void pnm_normalizetuple( struct pam * const pamP, tuple const tuple, tuplen const tuplen);

void pnm_unnormalizetuple( struct pam * const pamP, tuplen const tuplen, tuple const tuple);

void pnm_normalizeRow( struct pam * const pamP, const tuple * const tuplerow, pnm_transformMap * const transform, tuplen * const tuplenrow);

void pnm_unnormalizeRow( struct pam * const pamP, const tuplen * const tuplenrow, pnm_transformMap * const transform, tuple * const tuplerow);

void pnm_gammarown( struct pam * const pamP, tuplen * const row);

void pnm_ungammarown( struct pam * const pamP, tuplen * const row);

void pnm_applyopacityrown( struct pam * const pamP, tuplen * const tuplenrow);

void pnm_unapplyopacityrown( struct pam * const pamP, tuplen * const tuplenrow);

pnm_transformMap * pnm_creategammatransform( const struct pam * const pamP);

void pnm_freegammatransform( const pnm_transformMap * const transform, const struct pam * const pamP);

pnm_transformMap * pnm_createungammatransform( const struct pam * const pamP);

void pnm_freeungammatransform( const pnm_transformMap * const transform, const struct pam * const pamP);


pnm_YCbCrtuple() returns the Y/Cb/Cr luminance/chrominance representation of the color represented by the input tuple, assuming that the tuple is an RGB color representation (which is the case if it was read from a PPM image). The output components are based on the same scale (maxval) as the input tuple, but are floating point nonetheless to avoid losing information because of rounding. Divide them by the maxval to get normalized [0..1] values.

pnm_YCbCr_to_rgbtuple() does the reverse. pamP indicates the maxval for the returned tuple, and the Y, Cb, and Cr arguments are of the same scale.

It is possible for Y, Cb, and Cr to describe a color that cannot be represented in RGB form. In that case, pnm_YCbCr_to_rgbtuple() chooses a color as close as possible (by clipping each component to 0 and the maxval) and sets *overflowP true. It otherwise sets *overflowP false. pnm_lumin_factor[] is the factors (weights) one uses to compute the intensity of a color (according to some standard -- I don't know which). pnm_lumin_factor[0] is for the red component, [1] is for the green, and [2] is for the blue. They add up to 1.

pnm_gammarown() and pnm_ungammarown() apply and unapply gamma correction to a row of an image using the same transformation as pm_gamma709() and pm_ungamma709(). Note that these operate on a row of normalized tuples (tuplen, not tuple).

pnm_applyopacityrown() reduces the intensity of samples in accordance with the opacity plane of an image. The opacity plane, if it exists, tells how much of the light from that pixel should show when the image is composed with another image. You use pnm_applyopacityrown() in preparation for doing such a composition. For example, if the opacity plane says that the left half of the image is 50% opaque and the right half 100% opaque, pnm_applyopacityrown() will reduce the intensity of each sample of each tuple (pixel) in the left half of the image by 50%, and leave the rest alone.

If the image does not have an opacity plane (i.e. its tuple type is not one that libnetpbm recognizes as having an opacity plane), pnm_applyopacityrown() does nothing (which is the same as assuming opacity 100%). The tuple types that libnetpbm recognizes as having opacity are RGB_ALPHA and GRAYSCALE_ALPHA.

pnm_unapplyopacityrown() does the reverse. It assumes the intensities are already reduced according to the opacity plane, and raises back to normal.

pnm_applyopacityrown() works on (takes as input and produces as output) normalized, intensity-proportional tuples. That means you will typically read the row from the image file with pnm_readpamrown() and then gamma-correct it with pnm_ungammarown(), and then do pnm_applyopacityrown(). You then manipulate the row further (perhaps add it with other rows you've processed similarly), then do pnm_unapplyopacityrown(), then pnm_gammarown(), then pnm_writepamrown().

pnm_applyopacityrown() and pnm_unapplyopacityrown() were new in Netpbm 10.25 (October 2004).

pnm_normalizetuple() and pnm_unnormalizetuple() convert between a tuple data type and a tuplen data type. The former represents a sample value using the same unsigned integer that is in the PAM image, while the latter represents a sample value as a number scaled by the maxval to the range 0..1. I.e. pnm_normalizetuple() divides every sample value by the maxval and pnm_unnormalizetuple() multiples every sample by the maxval.

pnm_normalizeRow() and pnm_unnormalizeRow() do the same thing on an entire tuple row, but also have an extra feature: You can specify a transform function to be applied in addition. Typically, this is a gamma transform function. You can of course more easily apply your transform function separately from normalizing, but doing it all at once is usually way faster. Why? Because you can use a lookup table that is indexed by an integer on one side and produces a floating point number on the other. To do it separately, you'd either have to do floating point arithmetic on the normalized value or do the transform on the integer values and lose a lot of precision.

If you don't have any transformation to apply, just specify NULL for the transform argument and the function will just normalize (i.e. divide or multiply by the maxval).

Here's an example of doing a transformation. The example composes two images together, something that has to be done with intensity-linear sample values.

pnm_transformMap * const transform1 = pnm_createungammatransform(&inpam1);
pnm_transformMap * const transform2 = pnm_createungammatransform(&inpam2);
pnm_transformMap * const transformOut = pnm_creategammatransform(&outpam);

pnm_readpamrow(&inpam1, inrow1);
pnm_readpamrow(&inpam2, inrow2);

pnm_normalizeRow(&inpam1, inrow1, transform1, normInrow1);
pnm_normalizeRow(&inpam2, inrow2, transform2, normInrow2);

for (col = 0; col < outpam.width; ++col)
    normOutrow[col] = (normInrow1[col] + normInrow2[col])/2;

pnm_unnormalizeRow(&outpam, normOutrow, transformOut, outrow);

pnm_writepamrow(&outpam, outrow);

To specify a transform, you must create a special pnm_transformMap object and pass it as the transform argument. Typically, your transform is a gamma transformation because you want to work in intensity-proportional sample values and the PAM image format uses gamma-adjusted ones. In that case, just use pnm_creategammatransform() and pnm_createungammatransform() to create this object and don't worry about what's inside it.

pnm_creategammatransform() and pnm_createungammatransform() create objects that you use with pnm_normalizeRow() and pnm_unnormalizeRow() as described above. The created object describes a transform that applies or reverses the ITU-R Recommendation BT.709 gamma adjustment that is used in PAM visual images and normalizes or unnormalizes the sample values. pnm_freegammatransform() and pnm_freeungammatransform() destroy the objects.



void pnm_checkpam( struct pam *pamP, const enum pm_check_type check_type, enum pm_check_code *retvalP);

void pnm_nextimage( FILE *file, int * const eofP);


pnm_checkpam() checks for the common file integrity error where the file is the wrong size to contain the raster, according to the information in the header.

pnm_nextimage()positions a Netpbm image input file to the next image in it (so that a subsequent pnm_readpaminit() reads its header).

Table Of Contents