- experimenting with different signatures comparision algorithm

diff --git a/cf/images b/cf/images
index 9d767e6f7421e7a9ca0a83d6a09d09029e617f6b..09d148d183904c16ee8e151366b1c47593d33d23 100644 (file)
--- a/cf/images
+++ b/cf/images
@@ -19,6 +19,8 @@ PostQuantThreshold    2
 
 TexturedThreshold      0.32
 
+CompareMethod          2
+
 }
 
 ######## Duplicates finder ######################################################

diff --git a/images/config.c b/images/config.c
index a6e69a19886529c5a9b172e269256b63e5c60190..931cc1992f42c844b44a31f9f8e7e0fd35f6812d 100644 (file)
--- a/images/config.c
+++ b/images/config.c
@@ -20,6 +20,7 @@ uns image_sig_postquant_min_steps;
 uns image_sig_postquant_max_steps;
 uns image_sig_postquant_threshold;
 double image_sig_textured_threshold;
+uns image_sig_compare_method;
 
 static struct cf_section sig_config = {
   CF_ITEMS{
@@ -30,6 +31,7 @@ static struct cf_section sig_config = {
     CF_UNS("PostQuantMaxSteps", &image_sig_postquant_max_steps),
     CF_UNS("PostQuantThreshold", &image_sig_postquant_threshold),
     CF_DOUBLE("TexturedThreshold", &image_sig_textured_threshold),
+    CF_UNS("CompareMethod", &image_sig_compare_method),
     CF_END
   }
 };

diff --git a/images/sig-cmp.c b/images/sig-cmp.c
index ec3f1e865303beac7c058ef1bf262976e97493f7..c4137239608f435150761125b53490c7d8686e0d 100644 (file)
--- a/images/sig-cmp.c
+++ b/images/sig-cmp.c
@@ -14,12 +14,13 @@
 #include "images/math.h"
 #include "images/images.h"
 #include "images/signature.h"
+
 #include <stdio.h>
 
-uns
-image_signatures_dist(struct image_signature *sig1, struct image_signature *sig2)
+static uns
+image_signatures_dist_1(struct image_signature *sig1, struct image_signature *sig2)
 {
-  DBG("image_signatures_dist()");
+  DBG("image_signatures_dist_1()");
 
   uns cnt1 = sig1->len;
   uns cnt2 = sig2->len;
@@ -110,3 +111,109 @@ image_signatures_dist(struct image_signature *sig1, struct image_signature *sig2
 
   return (1 << (3 + 3 + 8 + 16)) - measure;
 }
+
+#define ASORT_PREFIX(x) image_signatures_dist_2_##x
+#define ASORT_KEY_TYPE uns
+#define ASORT_ELT(i) items[i]
+#define ASORT_EXTRA_ARGS , uns *items
+#include "lib/arraysort.h"
+
+static uns
+image_signatures_dist_2(struct image_signature *sig1, struct image_signature *sig2)
+{
+  DBG("image_signatures_dist_2()");
+
+  uns dist[IMAGE_REG_MAX * IMAGE_REG_MAX], p[IMAGE_REG_MAX], q[IMAGE_REG_MAX];
+  uns n, i, j, k, l, s, d;
+  struct image_region *reg1, *reg2;
+
+  /* Compute distance matrix */
+  n = 0;
+  /* ... for non-textured images */
+  if ((sig1->flags | sig2->flags) & IMAGE_SIG_TEXTURED)
+    for (j = 0, reg2 = sig2->reg; j < sig2->len; j++, reg2++)
+      for (i = 0, reg1 = sig1->reg; i < sig1->len; i++, reg1++)
+        {
+         // FIXME
+         /*uns ds =
+           isqr(reg1->h[0], reg2->h[0]) +
+           isqr(reg1->h[1], reg2->h[1]) +
+           isqr(reg1->h[2], reg2->h[2]);*/
+         uns dt =
+           isqr(reg1->f[0] - reg2->f[0]) +
+           isqr(reg1->f[1] - reg2->f[1]) +
+           isqr(reg1->f[2] - reg2->f[2]) +
+           isqr(reg1->f[3] - reg2->f[3]) +
+           isqr(reg1->f[4] - reg2->f[4]) +
+           isqr(reg1->f[5] - reg2->f[5]);
+         dist[n++] = (CLAMP(dt * 0xffff / (64 * 64 * 6), 0, 0xffff) << 8) + i + (j << 4) ;
+        }
+  /* ... for textured images (ignore shape properties) */
+  else
+    for (j = 0, reg2 = sig2->reg; j < sig2->len; j++, reg2++)
+      for (i = 0, reg1 = sig1->reg; i < sig1->len; i++, reg1++)
+        {
+         uns dt =
+           isqr(reg1->f[0] - reg2->f[0]) +
+           isqr(reg1->f[1] - reg2->f[1]) +
+           isqr(reg1->f[2] - reg2->f[2]) +
+           isqr(reg1->f[3] - reg2->f[3]) +
+           isqr(reg1->f[4] - reg2->f[4]) +
+           isqr(reg1->f[5] - reg2->f[5]);
+         dist[n++] = (CLAMP(dt * 0xffff / (64 * 64 * 6), 0, 0xffff) << 8) + i + (j << 4) ;
+        }
+
+  /* One or both signatures have no regions */
+  if (!n)
+    return 1 << IMAGE_SIG_DIST_SCALE;
+
+  /* Get percentages */
+  for (i = 0, reg1 = sig1->reg; i < sig1->len; i++, reg1++)
+    p[i] = reg1->wa;
+  for (i = 0, reg2 = sig2->reg; i < sig2->len; i++, reg2++)
+    q[i] = reg2->wa;
+
+  /* Sort entries in distance matrix */
+  image_signatures_dist_2_sort(n, dist);
+
+  /* Compute significance matrix and resulting distance */
+  uns sum = 0;
+  for (k = 0, l = 128; l; k++)
+    {
+      i = dist[k] & 15;
+      j = (dist[k] >> 4) & 15;
+      d = dist[k] >> 8;
+      if (p[i] <= q[j])
+        {
+         s = p[i];
+         q[j] -= p[i];
+         p[i] = 0;
+       }
+      else
+        {
+         s = q[j];
+         p[i] -= q[j];
+         q[j] = 0;
+       }
+      l -= s;
+      sum += s * d;
+      DBG("s[%u][%u]=%u d=%u", i, j, s, d);
+    }
+
+  return sum << (IMAGE_SIG_DIST_SCALE - 7 - 16);
+}
+
+uns
+image_signatures_dist(struct image_signature *sig1, struct image_signature *sig2)
+{
+  switch (image_sig_compare_method)
+    {
+      case 1:
+       return image_signatures_dist_1(sig1, sig2);
+      case 2:
+       return image_signatures_dist_2(sig1, sig2);
+      default:
+       die("Invalid image signatures compare method.");
+    }
+}
+

diff --git a/images/signature.h b/images/signature.h
index 1bf166aa3fba6b578649faab13320b84381802aa..d59e043d902824e67b17fdd5e47d6e902ddb89c5 100644 (file)
--- a/images/signature.h
+++ b/images/signature.h
@@ -6,6 +6,7 @@ extern uns image_sig_min_width, image_sig_min_height;
 extern uns *image_sig_prequant_thresholds;
 extern uns image_sig_postquant_min_steps, image_sig_postquant_max_steps, image_sig_postquant_threshold;
 extern double image_sig_textured_threshold;
+extern uns image_sig_compare_method;
 
 #define IMAGE_VEC_F    6
 #define IMAGE_REG_F    IMAGE_VEC_F
@@ -104,7 +105,7 @@ void image_sig_detect_textured(struct image_sig_data *data);
 
 /* sig-cmp.c */
 
-#define IMAGE_SIG_DIST_SCALE (3 + 3 + 8 + 16)
+#define IMAGE_SIG_DIST_SCALE 30
 
 uns image_signatures_dist(struct image_signature *sig1, struct image_signature *sig2);