OSDN Git Service

New cropping algorithm - determine cropping rectangle from a median filter run across...
[handbrake-jp/handbrake-jp-git.git] / libhb / scan.c
index d66ccca..f7ded05 100644 (file)
@@ -26,6 +26,21 @@ static int  DecodePreviews( hb_scan_t *, hb_title_t * title );
 static void LookForAudio( hb_title_t * title, hb_buffer_t * b );
 static int  AllAudioOK( hb_title_t * title );
 
+static const char *aspect_to_string( int aspect )
+{
+    switch ( aspect )
+    {
+        case HB_ASPECT_BASE * 1 / 1:    return "1:1";
+        case HB_ASPECT_BASE * 4 / 3:    return "4:3";
+        case HB_ASPECT_BASE * 16 / 9:   return "16:9";
+        case HB_ASPECT_BASE * 221 / 100:   return "2.21:1";
+    }
+    static char arstr[32];
+    double a = (double)aspect / HB_ASPECT_BASE;
+    sprintf( arstr, aspect >= 1.? "%.2f:1" : "1:%.2f", a );
+    return arstr;
+}
+
 hb_thread_t * hb_scan_init( hb_handle_t * handle, const char * path,
                             int title_index, hb_list_t * list_title )
 {
@@ -234,6 +249,106 @@ static void ScanFunc( void * _data )
     _data = NULL;
 }
 
+// -----------------------------------------------
+// stuff related to cropping
+
+#define DARK 64
+
+static int row_all_dark( hb_title_t *title, uint8_t* luma, int row )
+{
+    int i = title->width;
+    luma += i * row;
+    while ( --i >= 0 )
+    {
+        if ( *luma++ > DARK )
+            return 0;
+    }
+    return 1;
+}
+
+static int column_all_dark( hb_title_t *title, uint8_t* luma, int top, int col )
+{
+    int i = title->height - top;
+    int stride = title->width;
+    luma += stride * top + col;
+    for ( ; --i >= 0; luma += stride )
+    {
+        if ( *luma > DARK )
+            return 0;
+    }
+    return 1;
+}
+#undef DARK
+
+typedef struct {
+    int n;
+    int t[10];
+    int b[10];
+    int l[10];
+    int r[10];
+} crop_record_t;
+
+static void record_crop( crop_record_t *crops, int t, int b, int l, int r )
+{
+    crops->t[crops->n] = t;
+    crops->b[crops->n] = b;
+    crops->l[crops->n] = l;
+    crops->r[crops->n] = r;
+    ++crops->n;
+}
+
+static int compare_int( const void *a, const void *b )
+{
+    return *(const int *)a - *(const int *)b;
+}
+
+static void sort_crops( crop_record_t *crops )
+{
+    qsort( crops->t, crops->n, sizeof(crops->t[0]), compare_int );
+    qsort( crops->b, crops->n, sizeof(crops->t[0]), compare_int );
+    qsort( crops->l, crops->n, sizeof(crops->t[0]), compare_int );
+    qsort( crops->r, crops->n, sizeof(crops->t[0]), compare_int );
+}
+
+// -----------------------------------------------
+// stuff related to title width/height/aspect info
+
+typedef struct {
+    int count;              /* number of times we've seen this info entry */
+    hb_work_info_t info;    /* copy of info entry */
+} info_list_t;
+
+static void remember_info( info_list_t *info_list, hb_work_info_t *info )
+{
+    for ( ; info_list->count; ++info_list )
+    {
+        if ( memcmp( &info_list->info, info, sizeof(*info) ) == 0 )
+        {
+            // we found a match - bump its count
+            ++info_list->count;
+            return;
+        }
+    }
+    // no match found - add new entry to list (info_list points to
+    // the first free slot). NB - we assume that info_list was allocated
+    // so that it's big enough even if there are no dups. I.e., 10 slots
+    // allocated if there are 10 previews.
+    info_list->count = 1;
+    info_list->info = *info;
+}
+
+static void most_common_info( info_list_t *info_list, hb_work_info_t *info )
+{
+    int i, biggest = 0;
+    for ( i = 1; info_list[i].count; ++i )
+    {
+        if ( info_list[i].count > info_list[biggest].count )
+            biggest = i;
+    }
+    *info = info_list[biggest].info;
+    free( info_list );
+}
+
 /***********************************************************************
  * DecodePreviews
  ***********************************************************************
@@ -248,8 +363,8 @@ static int DecodePreviews( hb_scan_t * data, hb_title_t * title )
     hb_list_t     * list_es;
     int progressive_count = 0;
     int interlaced_preview_count = 0;
-    double last_ar = 0;
-    int ar16_count = 0, ar4_count = 0;
+    info_list_t * info_list = calloc( 10+1, sizeof(*info_list) );
+    crop_record_t *crops = calloc( 1, sizeof(*crops) );
 
     buf_ps   = hb_buffer_init( HB_DVD_READ_BUFFER_SIZE );
     list_es  = hb_list_init();
@@ -261,7 +376,7 @@ static int DecodePreviews( hb_scan_t * data, hb_title_t * title )
 
     for( i = 0; i < 10; i++ )
     {
-        int j, k;
+        int j;
         FILE * file_preview;
         char   filename[1024];
 
@@ -350,44 +465,12 @@ static int DecodePreviews( hb_scan_t * data, hb_title_t * title )
         vid_decoder->close( vid_decoder );
         free( vid_decoder );
 
+        remember_info( info_list, &vid_info );
+
         title->width = vid_info.width;
         title->height = vid_info.height;
-        title->pixel_aspect_width = vid_info.pixel_aspect_width;
-        title->pixel_aspect_height = vid_info.pixel_aspect_height;
-        
         title->rate = vid_info.rate;
         title->rate_base = vid_info.rate_base;
-        if ( vid_info.aspect != 0 )
-        {
-            if ( vid_info.aspect != last_ar && last_ar != 0 )
-            {
-                hb_log( "aspect ratio changed from %g to %g",
-                        last_ar, vid_info.aspect );
-            }
-            
-            if( !title->pixel_aspect_width && !title->pixel_aspect_height )
-            {
-                /* We don't have pixel aspect info from the source, so we're
-                   going to have to make a guess on the display aspect ratio. */
-                switch ( (int)vid_info.aspect )
-                {
-                    case HB_ASPECT_BASE * 4 / 3:
-                        ++ar4_count;
-                        break;
-                    case HB_ASPECT_BASE * 16 / 9:
-                        ++ar16_count;
-                        break;
-                    default:
-                        hb_log( "unknown aspect ratio %g", vid_info.aspect );
-                        /* if the aspect is closer to 4:3 use that
-                         * otherwise use 16:9 */
-                        vid_info.aspect < HB_ASPECT_BASE * 14 / 9 ? ++ar4_count :
-                                                                    ++ar16_count;
-                        break;
-                }
-            }
-            last_ar = vid_info.aspect;
-        }
 
         if( title->rate_base == 1126125 )
         {
@@ -427,13 +510,6 @@ static int DecodePreviews( hb_scan_t * data, hb_title_t * title )
             title->rate_base = 1126125;
         }
 
-        // start from third frame to skip opening logos
-        if( i == 2)
-        {
-            title->crop[0] = title->crop[1] = title->height / 2;
-            title->crop[2] = title->crop[3] = title->width / 2;
-        }
-
         while( ( buf_es = hb_list_item( list_es, 0 ) ) )
         {
             hb_list_rem( list_es, buf_es );
@@ -462,42 +538,52 @@ static int DecodePreviews( hb_scan_t * data, hb_title_t * title )
             hb_log( "scan: fopen failed (%s)", filename );
         }
 
-#define Y    vid_buf->data
-#define DARK 64
-
         /* Detect black borders */
 
-        for( j = 0; j < title->width; j++ )
+#define Y    vid_buf->data
+        int top, bottom, left, right;
+        int h4 = title->height / 4, w4 = title->width / 4;
+        for ( top = 2; top < h4; ++top )
         {
-            for( k = 2; k < title->crop[0]; k++ )
-                if( Y[ k * title->width + j ] > DARK )
-                {
-                    title->crop[0] = k;
-                    break;
-                }
-            for( k = 0; k < title->crop[1]; k++ )
-                if( Y[ ( title->height - k - 1 ) *
-                       title->width + j ] > DARK )
-                {
-                    title->crop[1] = k;
-                    break;
-                }
+            if ( ! row_all_dark( title, Y, top ) )
+                break;
         }
-        for( j = 0; j < title->height; j++ )
+        if ( top < 4 )
         {
-            for( k = 0; k < title->crop[2]; k++ )
-                if( Y[ j * title->width + k ] > DARK )
-                {
-                    title->crop[2] = k;
-                    break;
-                }
-            for( k = 0; k < title->crop[3]; k++ )
-                if( Y[ j * title->width +
-                        title->width - k - 1 ] > DARK )
-                {
-                    title->crop[3] = k;
-                    break;
-                }
+            // we started at row two to avoid the "line 19" noise that shows
+            // up on row 0 & 1 of some TV shows. Since we stopped before row 4
+            // check if row 0 & 1 are dark or if we shouldn't crop the top at all.
+            if ( row_all_dark( title, Y, 0 ) )
+            {
+                top = row_all_dark( title, Y, 1 )? top : 1;
+            }
+            else
+            {
+                top = 0;
+            }
+        }
+        for ( bottom = 0; bottom < h4; ++bottom )
+        {
+            if ( ! row_all_dark( title, Y, title->height - 1 - bottom ) )
+                break;
+        }
+        for ( left = 0; left < w4; ++left )
+        {
+            if ( ! column_all_dark( title, Y, top, left ) )
+                break;
+        }
+        for ( right = 0; right < w4; ++right )
+        {
+            if ( ! column_all_dark( title, Y, top, title->width - 1 - right ) )
+                break;
+        }
+
+        // only record the result if all the crops are less than a quarter of
+        // the frame otherwise we can get fooled by frames with a lot of black
+        // like titles, credits & fade-thru-black transitions.
+        if ( top < h4 && bottom < h4 && left < w4 && right < w4 )
+        {
+            record_crop( crops, top, bottom, left, right );
         }
         ++npreviews;
 
@@ -505,41 +591,66 @@ skip_preview:
         if ( vid_buf )
             hb_buffer_close( &vid_buf );
     }
-    
-    if( title->pixel_aspect_width && title->pixel_aspect_width )
-    {
-        title->aspect = ( (double)title->pixel_aspect_width * 
-                         (double)title->width /
-                         (double)title->pixel_aspect_height /
-                         (double)title->height + 0.05 ) * HB_ASPECT_BASE;
-    }
-    else
+
+    if ( npreviews )
     {
-        /* if we found mostly 4:3 previews use that as the aspect ratio otherwise
-           use 16:9 */
-        title->aspect = ar4_count > ar16_count ?
-                            HB_ASPECT_BASE * 4 / 3 : HB_ASPECT_BASE * 16 / 9;
-    }
+        // use the most common frame info for our final title dimensions
+        hb_work_info_t vid_info;
+        most_common_info( info_list, &vid_info );
 
-    title->crop[0] = EVEN( title->crop[0] );
-    title->crop[1] = EVEN( title->crop[1] );
-    title->crop[2] = EVEN( title->crop[2] );
-    title->crop[3] = EVEN( title->crop[3] );
+        title->width = vid_info.width;
+        title->height = vid_info.height;
+        title->pixel_aspect_width = vid_info.pixel_aspect_width;
+        title->pixel_aspect_height = vid_info.pixel_aspect_height;
 
-    hb_log( "scan: %d previews, %dx%d, %.3f fps, autocrop = %d/%d/%d/%d, aspect %.2f",
-            npreviews, title->width, title->height, (float) title->rate /
-            (float) title->rate_base, title->crop[0], title->crop[1],
-            title->crop[2], title->crop[3], (float)title->aspect/(float)HB_ASPECT_BASE);
+        // compute the aspect ratio based on the storage dimensions and the
+        // pixel aspect ratio (if supplied) or just storage dimensions if no PAR.
+        title->aspect = ( (double)title->width / (double)title->height + 0.05 ) *
+                        HB_ASPECT_BASE;
 
-    if( interlaced_preview_count >= ( npreviews / 2 ) )
-    {
-        hb_log("Title is likely interlaced or telecined (%i out of %i previews). You should do something about that.",
-               interlaced_preview_count, npreviews);
-        title->detected_interlacing = 1;
-    }
-    else
-    {
-        title->detected_interlacing = 0;
+        double aspect = (double)title->width / (double)title->height;
+        if( title->pixel_aspect_width && title->pixel_aspect_height )
+        {
+            aspect *= (double)title->pixel_aspect_width /
+                      (double)title->pixel_aspect_height;
+        }
+        title->aspect = ( aspect + 0.05 ) * HB_ASPECT_BASE;
+
+        if ( crops->n )
+        {
+            sort_crops( crops );
+            // The next line selects median cropping - at least
+            // 50% of the frames will have their borders removed.
+            // Other possible choices are loose cropping (i = 0) where 
+            // no non-black pixels will be cropped from any frame and a
+            // tight cropping (i = crops->n - (crops->n >> 2)) where at
+            // least 75% of the frames will have their borders removed.
+            i = crops->n >> 1;
+            title->crop[0] = EVEN( crops->t[i] );
+            title->crop[1] = EVEN( crops->b[i] );
+            title->crop[2] = EVEN( crops->l[i] );
+            title->crop[3] = EVEN( crops->r[i] );
+        }
+        free( crops );
+
+        hb_log( "scan: %d previews, %dx%d, %.3f fps, autocrop = %d/%d/%d/%d, "
+                "aspect %s, PAR %d:%d",
+                npreviews, title->width, title->height, (float) title->rate /
+                (float) title->rate_base,
+                title->crop[0], title->crop[1], title->crop[2], title->crop[3],
+                aspect_to_string( title->aspect ), title->pixel_aspect_width,
+                title->pixel_aspect_height );
+
+        if( interlaced_preview_count >= ( npreviews / 2 ) )
+        {
+            hb_log("Title is likely interlaced or telecined (%i out of %i previews). You should do something about that.",
+                   interlaced_preview_count, npreviews);
+            title->detected_interlacing = 1;
+        }
+        else
+        {
+            title->detected_interlacing = 0;
+        }
     }
 
     hb_buffer_close( &buf_ps );