path: root/src/itrans-stat.c

/*
 * itrans-stat.c
 *
 * Peak detection by iterative statistical analysis and processing
 *
 * (c) 2007 Gordon Ball <gfb21@cam.ac.uk>
 *	    Thomas White <taw27@cam.ac.uk>
 *
 *  dtr - Diffraction Tomography Reconstruction
 *
 */

#ifdef HAVE_CONFIG_H
#include <config.h>
#endif

#include <math.h>
#include <gsl/gsl_matrix.h>

#include "reflections.h"
#include "control.h"
#include "imagedisplay.h"
#include "utils.h"

/* 
 * Create a gsl_matrix for performing image operations on
 * from a raw image and control context
 * Converting to gsl_matrix because many of the required operations
 * can be done as matrices to save effort
 * Renormalises matrix to 0->1
 */
static gsl_matrix *itrans_peaksearch_stat_createImageMatrix(ControlContext *ctx, int16_t *image) {

	gsl_matrix *raw;
	int i, j;

	raw = gsl_matrix_alloc(ctx->width,ctx->height);
	for ( i=0; i<raw->size1; i++ ) {
		for ( j=0; j<raw->size2; j++ ) {
			gsl_matrix_set(raw, i, j, (double)image[i+j*ctx->width]);
		}
	}
	matrix_renormalise(raw);
	
	return raw;

}

/* Find and return the mean value of the matrix */
static double itrans_peaksearch_stat_image_mean(gsl_matrix *m) {

	int i, j;
	double mean = 0.;

	for ( i=0; i<m->size1; i++ ) {
		for ( j=0; j<m->size2; j++ ) {
			mean += gsl_matrix_get(m,i,j);
		}
	}

	return mean / (m->size1 * m->size2);

}

/*
 * Return the standard deviation, sigma, of the matrix values
 * \sqrt(\sum((x-mean)^2)/n)
 */
static double itrans_peaksearch_stat_image_sigma(gsl_matrix *m, double mean) {

	int i,j;
	double diff2 = 0;

	for ( i=0; i<m->size1; i++ ) {
		for ( j=0; j<m->size2; j++ ) {
			diff2 += (gsl_matrix_get(m,i,j)-mean)*(gsl_matrix_get(m,i,j)-mean);
		}
	}
	
	return sqrt(diff2/(m->size1 * m->size2));

}

/*
 * Filter all pixels with value < mean + k*sigma to 0
 * Set all matching pixels to 1 d
 */
static void itrans_peaksearch_stat_sigma_filter(gsl_matrix *m, double k) {

	double mean,sigma;
	int i,j;

	mean = itrans_peaksearch_stat_image_mean(m);
	sigma = itrans_peaksearch_stat_image_sigma(m,mean);

	for ( i=0; i<m->size1; i++ ) {
		for ( j=0; j<m->size2; j++ ) {
			if (gsl_matrix_get(m,i,j) >= mean+k*sigma) {
				gsl_matrix_set(m, i, j, 1.);
			} else {
				gsl_matrix_set(m, i, j, 0.);
			}
		}
	}
	
}

/*
 * Calculate the mean within a circle centred (x,y) of radius r
 *
 * TODO: Use a mask instead of calculating valid points
 */
static double itrans_peaksearch_stat_circle_mean(gsl_matrix *m, int x, int y, int r, gsl_matrix *mask) {

	double mean = 0.;
	int i, j;
	int n = 0;

	for ( i=x-r; i<=x+r; i++ ) {
		for ( j=y-r; j<=y+r; j++ ) {
			//printf("cm: ij=(%d,%d) mask=(%d,%d)\n",i,j,i-x+r,j-y+r);
			if ( gsl_matrix_get(mask,i-x+r,j-y+r)>0. ) {
				mean += gsl_matrix_get(m, i, j);
				//printf("cm: (%d,%d) mean=%lf val=%lf\n",i,j,mean,gsl_matrix_get(m,i,j));
				n++;
			}
		}
	}
	
	//printf("cm: (%d,%d) summean=%lf n=%d\n",x,y,mean,n);
	return mean/n;
	
}

/*
 * Calculate sigma within a circle centred (x,y) of radius r
 *
 * TODO: Use a mask instead of calculating valid points
 */
static double itrans_peaksearch_stat_circle_sigma(gsl_matrix *m, int x, int y, int r, gsl_matrix *mask, double mean) {

	double diff2 = 0.;
	int i, j;
	int n = 0;

	for ( i=x-r; i<=x+r; i++ ) {
		for ( j=y-r; j<=y+r; j++ ) {
			if ( gsl_matrix_get(mask, i-x+r, j-y+r) > 0 ) {
				diff2 += (gsl_matrix_get(m,i,j)-mean)*(gsl_matrix_get(m,i,j)-mean);
				n++;
			}
		}
	}
	
	return sqrt(diff2/n);

}

/*
 * Calculate a circular mask to save recalculating it every time
 */
static gsl_matrix *itrans_peaksearch_stat_circle_mask(int r) {

	gsl_matrix *m;
	int i,j;

	m = gsl_matrix_calloc(2*r+1, 2*r+1);
	for ( i=0; i<2*r+1; i++ ) {
		for ( j=0; j<2*r+1; j++ ) {
			if ( sqrt((r-i)*(r-i)+(r-j)*(r-j)) <= r ) {
				 gsl_matrix_set(m, i, j, 1.);
			}
			
		}
	}
	
	return m;
	
}

/*
 * Variation on the above filter where instead of using
 * sigma and mean for the whole image, it is found for a local
 * circle of radius r pixels.
 * The central point is also calculated as an approximately gaussian
 * average over local pixels rather than just a single pixel.
 * This takes a long time - 20-30 seconds for a 1024^2 image and r=10,
 * obviously large values of r will make it even slower.
 * The appropriate r value needs to be considered carefully - it needs to
 * be somewhat smaller than the average inter-reflection seperation.
 * 10 pixels worked well for the sapphire.mrc images, but this might
 * not be the case for other images. Images for which this would be very
 * large probably need to be resampled to a lower resolution.
 *
 * TODO: Pass a mask to the ancilliary functions instead of having
 * them calculate several hundred million sqrts
 * 
 * self-referencing problem being dealt with - output being written onto the array before the next point it computed
 * problem carried over from the OO version where a new object was created by each operation
 */
static void itrans_peaksearch_stat_local_sigma_filter(gsl_matrix *m, int r, double k) {

	double mean,sigma;
	double local;
	gsl_matrix *mask;
	gsl_matrix *new;
	int i,j;
	//int interval;

	mask = itrans_peaksearch_stat_circle_mask(r);
	new = gsl_matrix_alloc(m->size1, m->size2);

	//interval = (m->size1-r)/20;
	//printf("lsf: starting loop\n");
	//printf("lsf: ");
	//for (i=r;i<m->size1-r;i++) {
	//	for (j=r;j<m->size2-r;j++) {

	for ( i=0; i<m->size1; i++ ) {
		for ( j=0; j<m->size2; j++ ) {
			if ( ((i >= r) && (i < m->size1-r)) && ((j >= r) && (j < m->size2-r)) ) {
			
				//printf("lsf: evaluating (%d,%d)\n",i,j);
				mean = itrans_peaksearch_stat_circle_mean(m, i, j, r, mask);
				//printf("lsf: mean=%lf",mean);
				sigma = itrans_peaksearch_stat_circle_sigma(m, i, j, r, mask, mean);
				//printf(" sigma=%lf",sigma);
				local = gsl_matrix_get(m, i, j);
				local += gsl_matrix_get(m, i+1, j) + gsl_matrix_get(m, i-1, j) + gsl_matrix_get(m, i, j+1) + gsl_matrix_get(m, i, j-1);
				local += .5*(gsl_matrix_get(m, i+1, j+1) + gsl_matrix_get(m, i-1, j+1) + gsl_matrix_get(m, i+1, j-1) + gsl_matrix_get(m, i-1, j-1));
				local /= 7.;
				//printf(" local=%lf\n",local);
				if ( local > mean+k*sigma ) {
					gsl_matrix_set(new, i, j, 1.);
				} else {
					gsl_matrix_set(new, i, j, 0.);
				}
				
			} else {
				gsl_matrix_set(new, i, j, 0.);
			}
		}
		//if (i % interval == 0) printf(".");
	}
	
	//printf("done\n");
	gsl_matrix_memcpy(m, new);
	gsl_matrix_free(new);

}

/*
 * Apply an arbitary kernel to the image - each point takes the value
 * of the sum of the kernel convolved with the surrounding pixels.
 * The kernel needs to be a (2n+1)^2 array (centred on (n,n)). It needs
 * to be square and should probably be normalised.
 * Don't do daft things like rectangular kernels or kernels larger
 * than the image - this doesn't check such things and will cry.
 * 
 * Also suffers from self-reference problem
 */
static void itrans_peaksearch_stat_apply_kernel(gsl_matrix *m, gsl_matrix *kernel) {

	int size;
	int half;
	gsl_matrix *l;
	gsl_matrix_view lv;
	gsl_matrix *new;
	double val;
	int i, j, x, y;

	size = kernel->size1;
	half = (size-1)/2;
	new = gsl_matrix_calloc(m->size1, m->size2);

	for ( i=0; i<m->size1; i++ ) {
		for ( j=0; j<m->size2; j++ ) {
			if ( ((i >= half) && (i < m->size1-half)) && ((j >= half) && (j < m->size2-half)) ) {
			
				lv = gsl_matrix_submatrix(m, i-half, j-half, size, size);
				l = &lv.matrix;
				val = 0.;
				for ( x=0; x<size; x++ ) {
					for ( y=0; y<size; y++ ) {
						val += gsl_matrix_get(l, x, y)*gsl_matrix_get(kernel, x, y);
					}
				}
				//gsl_matrix_free(l);
				gsl_matrix_set(new,i,j,val);
				
			}
		}
	}
	
	gsl_matrix_memcpy(m,new);
	gsl_matrix_free(new);
	
}

/* Generate the simplist possible kernel - a flat one */
static gsl_matrix *itrans_peaksearch_stat_generate_flat_kernel(int half) {

	gsl_matrix *k;

	k = gsl_matrix_alloc(2*half+1,2*half+1);
	gsl_matrix_set_all(k,1./((2*half+1)*(2*half+1)));

	return k;

}

/* Expands or contracts a gsl_matrix by copying the columns to a new one */
static gsl_matrix *itrans_peaksearch_stat_matrix_expand(gsl_matrix *m, int oldsize, int newsize) {

	gsl_matrix *new;
	int j;
	
	//printf("me: %d->%d\n",oldsize,newsize);
	
	new = gsl_matrix_calloc(2, newsize);
	//printf("me: calloc(2,%d)\n",newsize);
	for ( j=0; j<oldsize; j++) {
		if ( j < newsize ) {
			//printf("me: copying col %d\n",j);
			gsl_matrix_set(new, 0, j, gsl_matrix_get(m, 0, j));
			gsl_matrix_set(new, 1, j, gsl_matrix_get(m, 1, j));
		}
	}

	//printf("me: freeing old matrix\n");
	gsl_matrix_free(m);

	//printf("me: new s1=%d s2=%d\n",new->size1,new->size2);
	//printf("me: m s1=%d s2=%d\n",m->size1,m->size2);
	return new;
	
}

/*
 * Stack-based flood-fill iteration routine 
 * have to return a pointer to com each time because if the matrix size has to be changed then we need to keep
 * track of the location of the resized instance
 */
static gsl_matrix *itrans_peaksearch_stat_do_ff(int i, int j, int* mask, double threshold, gsl_matrix *m, gsl_matrix *com, int *com_n, int *com_size) {

	if ( (i >= 0) && (i < m->size1) ) {
		if ( (j >= 0) && (j < m->size2) ) {
			if ( mask[i+j*m->size1] == 0 ) {
				if ( gsl_matrix_get(m, i, j) > threshold ) {
				
					//printf("dff: found valid point (%d,%d)\n",i,j);
					gsl_matrix_set(com, 0, *com_n, i);
					gsl_matrix_set(com, 1, *com_n, j);
					*com_n = *com_n + 1;
					if ( *com_n == *com_size ) {
						com = itrans_peaksearch_stat_matrix_expand(com, *com_size, *com_size*2);
						*com_size *= 2;
					}
					mask[i+j*m->size1] = 1;
					com = itrans_peaksearch_stat_do_ff(i+1, j, mask,threshold, m, com, com_n, com_size);
					com = itrans_peaksearch_stat_do_ff(i-1, j, mask,threshold, m, com, com_n, com_size);
					com = itrans_peaksearch_stat_do_ff(i, j+1, mask,threshold, m, com, com_n, com_size);
					com = itrans_peaksearch_stat_do_ff(i, j-1, mask,threshold, m, com, com_n, com_size);
					
				}
			}
		}
	}
	
	return com;

}

/*
 * Find points by floodfilling all pixels above threshold
 * Implements a stack-based flood-filling method which may
 * cause stack overflows if the blocks are extremely large -
 * dependent on implementation (default 4MiB stack for unix?)
 * Implements a crude variable-sized array method which hopefully works
 * Returns a gsl_matrix with x coordinates in row 0 and y
 * coordinates in row 1, which should be of the right length
 * Variable count is set to the number of points found
 */
static gsl_matrix *itrans_peaksearch_stat_floodfill(gsl_matrix *m, double threshold, int *count) {

	int *mask;
	int size, com_size, i, j, k, n;
	int com_n;
	gsl_matrix *p;
	gsl_matrix *com;
	double com_x, com_y;

	mask = calloc(m->size1*m->size2, sizeof(int));
	size = 32;
	n = 0;
	p = gsl_matrix_calloc(2, size);

	//printf("ff: starting loop\n");
	for ( i=0; i<m->size1; i++ ) {
		for ( j=0; j<m->size2; j++ ) {
			if ( gsl_matrix_get(m, i, j) > threshold ) {
				if ( mask[i+j*m->size1] == 0 ) {
				
					//printf("ff: found starting point (%d,%d)\n",i,j);
					com_size = 32;
					com_n = 0;
					com_x = com_y = 0.;					
					com = gsl_matrix_calloc(2, com_size); //this is going to hold the points found for this location
					//printf("ff: starting floodfill stack\n");
					com = itrans_peaksearch_stat_do_ff(i, j, mask, threshold, m, com, &com_n, &com_size);
					//printf("ff: ended floodfill stack\n");
					for ( k=0; k<com_n; k++ ) {
						com_x += gsl_matrix_get(com, 0, k);
						com_y += gsl_matrix_get(com, 1, k);
					}
					com_x /= com_n;
					com_y /= com_n;
					//printf("ff: point CoM (%lf,%lf)\n",com_x,com_y);
					gsl_matrix_set(p, 0, n, com_x);
					gsl_matrix_set(p, 1, n, com_y);
					n++;
					if ( n == size ) {
						p = itrans_peaksearch_stat_matrix_expand(p, size, size*2);
						size *= 2;
					}
					
				}
			}
		}
	}
	//printf("ff: ending loop, found %d\n",n);

	*count = n;
	//printf("pcheck s1=%d s2=%d\n",p->size1,p->size2);
	p = itrans_peaksearch_stat_matrix_expand(p, size, n);
	//printf("pcheck s1=%d s2=%d\n",p->size1,p->size2);
	
	return p;
	
}

/* Implements the iteration based automatic method
 * returns a gsl_matrix formatted as described in flood-fill */
static gsl_matrix *itrans_peaksearch_stat_iterate(gsl_matrix *m, unsigned int *count) {

	int old;
	int cur;
	double k;
	double mean,sigma;
	gsl_matrix *p;
	gsl_matrix *kernel;
	
	old = m->size1*m->size2;
	
	//printf("Iterate: starting\n");
	//printf("Iterate: generating kernel\n");
	kernel = itrans_peaksearch_stat_generate_flat_kernel(1);
	//printf("Iterate: performing local_sigma_filter\n");
	itrans_peaksearch_stat_local_sigma_filter(m, 10, 1.);
	
	//printf("Iterate: starting loop\n");
	while ( 1 ) {
	
		//printf("Iterate: smoothing");
		itrans_peaksearch_stat_apply_kernel(m, kernel);
		//printf(" (1)");
		itrans_peaksearch_stat_apply_kernel(m, kernel);
		//printf(" (2)\n");
		mean = itrans_peaksearch_stat_image_mean(m);
		sigma = itrans_peaksearch_stat_image_sigma(m,mean);
		//printf("Iterate: mean=%lf sigma=%lf\n",mean,sigma);
		k = (0.5-mean)/sigma;
		//printf("Iterate: applying sigma_filter(%lf)\n",k);
		itrans_peaksearch_stat_sigma_filter(m,k);
		//printf("Iterate: floodfilling\n");
		p = itrans_peaksearch_stat_floodfill(m, 0, &cur);
		//printf("Iterate: %d points found\n", cur);
		if ( old < 1.05*cur ) break;
		old = cur;
		
	}
	
	gsl_matrix_free(kernel); 
	//printf("Iterate: finished\n");
	*count = cur;
	
	return p;

}

unsigned int itrans_peaksearch_stat(int16_t *image, ControlContext *ctx, double tilt_degrees, ImageDisplay *imagedisplay) {

	unsigned int n_reflections;	
	gsl_matrix *m;
	gsl_matrix *p;
	int i;
	double px,py;

	m = itrans_peaksearch_stat_createImageMatrix(ctx, image);
	p = itrans_peaksearch_stat_iterate(m, &n_reflections);
	
	for ( i=0; i<n_reflections; i++ ) {
	
		px = gsl_matrix_get(p,0,i);
		py = gsl_matrix_get(p,1,i);
		if ( ctx->fmode == FORMULATION_PIXELSIZE ) {
			reflection_add_from_reciprocal(ctx, (px-ctx->x_centre)*ctx->pixel_size, (py-ctx->y_centre)*ctx->pixel_size,
								tilt_degrees, 1.0);
		} else {
			reflection_add_from_dp(ctx, (px-ctx->x_centre), (py-ctx->y_centre), tilt_degrees, 1.0);
		}
		if (ctx->first_image) imagedisplay_mark_point(imagedisplay, (unsigned int)px, (unsigned int)py);
		
	}
	gsl_matrix_free(m);
	gsl_matrix_free(p);
	
	return n_reflections;
	
}