X-Git-Url: http://mj.ucw.cz/gitweb/?a=blobdiff_plain;f=images%2Fsig-init.c;h=3c76920fb36fc79d103b3b202270847332eb6389;hb=0db6e10eac28f38bfc3b325b13ad95107c58ce1e;hp=6b99e34309a51253582bce21b621c7aa25bdec65;hpb=956f5794fc8dad3ed98095f578cddbfe342f8db2;p=libucw.git diff --git a/images/sig-init.c b/images/sig-init.c index 6b99e343..3c76920f 100644 --- a/images/sig-init.c +++ b/images/sig-init.c @@ -9,72 +9,78 @@ #undef LOCAL_DEBUG -#include "sherlock/sherlock.h" -#include "lib/math.h" -#include "lib/fastbuf.h" -#include "lib/conf.h" -#include "lib/heap.h" -#include "images/math.h" -#include "images/images.h" -#include "images/color.h" -#include "images/signature.h" +#include +#include +#include +#include +#include +#include +#include +#include #include - -static double image_sig_inertia_scale[3] = { 3, 1, 0.3 }; - -struct block { - u32 area; /* block area in pixels (usually 16) */ - u32 v[IMAGE_VEC_F]; - u32 x, y; /* block position */ - struct block *next; -}; +#include int -compute_image_signature(struct image_thread *thread UNUSED, struct image_signature *sig, struct image *image) +image_sig_init(struct image_context *ctx, struct image_sig_data *data, struct image *image) { - bzero(sig, sizeof(*sig)); ASSERT((image->flags & IMAGE_PIXEL_FORMAT) == COLOR_SPACE_RGB); - uns cols = image->cols; - uns rows = image->rows; - uns row_size = image->row_size; - - uns w = (cols + 3) >> 2; - uns h = (rows + 3) >> 2; - - DBG("Computing signature for image of %ux%u pixels (%ux%u blocks)", cols, rows, w, h); + data->image = image; + data->flags = 0; + data->cols = (image->cols + 3) >> 2; + data->rows = (image->rows + 3) >> 2; + data->full_cols = image->cols >> 2; + data->full_rows = image->rows >> 2; + data->blocks_count = data->cols * data->rows; + if (data->blocks_count >= 0x10000) + { + IMAGE_ERROR(ctx, IMAGE_ERROR_INVALID_DIMENSIONS, "Image too large for implemented signature algorithm."); + return 0; + } + data->blocks = xmalloc(data->blocks_count * sizeof(struct image_sig_block)); + data->area = image->cols * image->rows; + DBG("Computing signature for image of %ux%u pixels (%ux%u blocks)", + image->cols, image->rows, data->cols, data->rows); + return 1; +} - uns blocks_count = w * h; - struct image_sig_block *blocks = xmalloc(blocks_count * sizeof(struct image_sig_block)), *block = blocks; +void +image_sig_preprocess(struct image_sig_data *data) +{ + struct image *image = data->image; + struct image_sig_block *block = data->blocks; + uns sum[IMAGE_VEC_F]; + bzero(sum, sizeof(sum)); /* Every block of 4x4 pixels */ byte *row_start = image->pixels; - for (uns block_y = 0; block_y < h; block_y++, row_start += row_size * 4) + for (uns block_y = 0; block_y < data->rows; block_y++, row_start += image->row_size * 4) { byte *p = row_start; - for (uns block_x = 0; block_x < w; block_x++, p += 12, block++) + for (uns block_x = 0; block_x < data->cols; block_x++, p += 12, block++) { int t[16], s[16], *tp = t; block->x = block_x; block->y = block_y; /* Convert pixels to Luv color space and compute average coefficients */ - uns l_sum = 0; - uns u_sum = 0; - uns v_sum = 0; + uns l_sum = 0, u_sum = 0, v_sum = 0; byte *p2 = p; - if ((!(cols & 3) || block_x < w - 1) && (!(rows & 3) || block_y < h - 1)) + if (block_x < data->full_cols && block_y < data->full_rows) { - for (uns y = 0; y < 4; y++, p2 += row_size - 12) + for (uns y = 0; y < 4; y++, p2 += image->row_size - 12) for (uns x = 0; x < 4; x++, p2 += 3) { byte luv[3]; srgb_to_luv_pixel(luv, p2); - l_sum += *tp++ = luv[0]; + l_sum += *tp++ = luv[0] / 4; u_sum += luv[1]; v_sum += luv[2]; } block->area = 16; + sum[0] += l_sum; + sum[1] += u_sum; + sum[2] += v_sum; block->v[0] = (l_sum >> 4); block->v[1] = (u_sum >> 4); block->v[2] = (v_sum >> 4); @@ -83,36 +89,39 @@ compute_image_signature(struct image_thread *thread UNUSED, struct image_signatu else { uns x, y; - uns square_cols = (block_x < w - 1 || !(cols & 3)) ? 4 : cols & 3; - uns square_rows = (block_y < h - 1 || !(rows & 3)) ? 4 : rows & 3; - for (y = 0; y < square_rows; y++, p2 += row_size) + uns square_cols = (block_x < data->full_cols) ? 4 : image->cols & 3; + uns square_rows = (block_y < data->full_rows) ? 4 : image->rows & 3; + for (y = 0; y < square_rows; y++, p2 += image->row_size) { byte *p3 = p2; for (x = 0; x < square_cols; x++, p3 += 3) { byte luv[3]; srgb_to_luv_pixel(luv, p3); - l_sum += *tp++ = luv[0]; + l_sum += *tp++ = luv[0] / 4; u_sum += luv[1]; v_sum += luv[2]; } for (; x < 4; x++) { - *tp = tp[-square_cols]; + *tp = tp[-(int)square_cols]; tp++; } } for (; y < 4; y++) for (x = 0; x < 4; x++) { - *tp = tp[-square_rows * 4]; + *tp = tp[-(int)square_rows * 4]; tp++; } block->area = square_cols * square_rows; - uns div = 0x10000 / block->area; - block->v[0] = (l_sum * div) >> 16; - block->v[1] = (u_sum * div) >> 16; - block->v[2] = (v_sum * div) >> 16; + uns inv = 0x10000 / block->area; + sum[0] += l_sum; + sum[1] += u_sum; + sum[2] += v_sum; + block->v[0] = (l_sum * inv) >> 16; + block->v[1] = (u_sum * inv) >> 16; + block->v[2] = (v_sum * inv) >> 16; } /* Apply Daubechies wavelet transformation */ @@ -135,65 +144,58 @@ compute_image_signature(struct image_thread *thread UNUSED, struct image_signatu /* ... and to the columns... skip LL band */ for (i = 0; i < 2; i++) { - t[i + 8] = (DAUB_3 * s[i + 8] - DAUB_2 * s[i +12] + DAUB_1 * s[i + 0] - DAUB_0 * s[i + 4]) / 0x2000; - t[i +12] = (DAUB_3 * s[i + 0] - DAUB_2 * s[i + 4] + DAUB_1 * s[i + 8] - DAUB_0 * s[i +12]) / 0x2000; + t[i + 8] = (DAUB_3 * s[i + 8] - DAUB_2 * s[i +12] + DAUB_1 * s[i + 0] - DAUB_0 * s[i + 4]) / 0x10000; + t[i +12] = (DAUB_3 * s[i + 0] - DAUB_2 * s[i + 4] + DAUB_1 * s[i + 8] - DAUB_0 * s[i +12]) / 0x10000; } for (; i < 4; i++) { - t[i + 0] = (DAUB_0 * s[i + 8] + DAUB_1 * s[i +12] + DAUB_2 * s[i + 0] + DAUB_3 * s[i + 4]) / 0x2000; - t[i + 4] = (DAUB_0 * s[i + 0] + DAUB_1 * s[i + 4] + DAUB_2 * s[i + 8] + DAUB_3 * s[i +12]) / 0x2000; - t[i + 8] = (DAUB_3 * s[i + 8] - DAUB_2 * s[i +12] + DAUB_1 * s[i + 0] - DAUB_0 * s[i + 4]) / 0x2000; - t[i +12] = (DAUB_3 * s[i + 0] - DAUB_2 * s[i + 4] + DAUB_1 * s[i + 8] - DAUB_0 * s[i +12]) / 0x2000; + t[i + 0] = (DAUB_0 * s[i + 8] + DAUB_1 * s[i +12] + DAUB_2 * s[i + 0] + DAUB_3 * s[i + 4]) / 0x10000; + t[i + 4] = (DAUB_0 * s[i + 0] + DAUB_1 * s[i + 4] + DAUB_2 * s[i + 8] + DAUB_3 * s[i +12]) / 0x10000; + t[i + 8] = (DAUB_3 * s[i + 8] - DAUB_2 * s[i +12] + DAUB_1 * s[i + 0] - DAUB_0 * s[i + 4]) / 0x10000; + t[i +12] = (DAUB_3 * s[i + 0] - DAUB_2 * s[i + 4] + DAUB_1 * s[i + 8] - DAUB_0 * s[i +12]) / 0x10000; } /* Extract energies in LH, HL and HH bands */ - block->v[3] = fast_sqrt_u16(isqr(t[8]) + isqr(t[9]) + isqr(t[12]) + isqr(t[13])); - block->v[4] = fast_sqrt_u16(isqr(t[2]) + isqr(t[3]) + isqr(t[6]) + isqr(t[7])); - block->v[5] = fast_sqrt_u16(isqr(t[10]) + isqr(t[11]) + isqr(t[14]) + isqr(t[15])); + block->v[3] = fast_sqrt_u32(isqr(t[8]) + isqr(t[9]) + isqr(t[12]) + isqr(t[13])); + block->v[4] = fast_sqrt_u32(isqr(t[2]) + isqr(t[3]) + isqr(t[6]) + isqr(t[7])); + block->v[5] = fast_sqrt_u32(isqr(t[10]) + isqr(t[11]) + isqr(t[14]) + isqr(t[15])); + sum[3] += block->v[3] * block->area; + sum[4] += block->v[4] * block->area; + sum[5] += block->v[5] * block->area; } } - /* FIXME: simple average is for testing pusposes only */ - uns l_sum = 0; - uns u_sum = 0; - uns v_sum = 0; - uns lh_sum = 0; - uns hl_sum = 0; - uns hh_sum = 0; - for (uns i = 0; i < blocks_count; i++) - { - l_sum += blocks[i].v[0]; - u_sum += blocks[i].v[1]; - v_sum += blocks[i].v[2]; - lh_sum += blocks[i].v[3]; - hl_sum += blocks[i].v[4]; - hh_sum += blocks[i].v[5]; - } - - sig->vec.f[0] = l_sum / blocks_count; - sig->vec.f[1] = u_sum / blocks_count; - sig->vec.f[2] = v_sum / blocks_count; - sig->vec.f[3] = lh_sum / blocks_count; - sig->vec.f[4] = hl_sum / blocks_count; - sig->vec.f[5] = hh_sum / blocks_count; + /* Compute featrures average */ + uns inv = 0xffffffffU / data->area; + for (uns i = 0; i < IMAGE_VEC_F; i++) + data->f[i] = ((u64)sum[i] * inv) >> 32; - if (cols < image_sig_min_width || rows < image_sig_min_height) + if (image->cols < image_sig_min_width || image->rows < image_sig_min_height) { - xfree(blocks); - return 1; + data->valid = 0; + data->regions_count = 0; } + else + data->valid = 1; +} - /* Quantize blocks to image regions */ - struct image_sig_region regions[IMAGE_REG_MAX]; - sig->len = image_sig_segmentation(blocks, blocks_count, regions); +void +image_sig_finish(struct image_sig_data *data, struct image_signature *sig) +{ + for (uns i = 0; i < IMAGE_VEC_F; i++) + sig->vec.f[i] = data->f[i]; + sig->len = data->regions_count; + sig->flags = data->flags; + if (!sig->len) + return; /* For each region */ u64 w_total = 0; - uns w_border = (MIN(w, h) + 3) / 4; - uns w_mul = 127 * 256 / w_border; + uns w_border = MIN(data->cols, data->rows) * image_sig_border_size; + int w_mul = w_border ? image_sig_border_bonus * 256 / (int)w_border : 0; for (uns i = 0; i < sig->len; i++) { - struct image_sig_region *r = regions + i; + struct image_sig_region *r = data->regions + i; DBG("Processing region %u: count=%u", i, r->count); ASSERT(r->count); @@ -206,40 +208,41 @@ compute_image_signature(struct image_thread *thread UNUSED, struct image_signatu sig->reg[i].f[5] = r->a[5]; /* Compute coordinates centroid and region weight */ - u64 x_avg = 0, y_avg = 0, w_sum = 0; + u64 x_sum = 0, y_sum = 0, w_sum = 0; for (struct image_sig_block *b = r->blocks; b; b = b->next) { - x_avg += b->x; - y_avg += b->y; + x_sum += b->x; + y_sum += b->y; uns d = b->x; d = MIN(d, b->y); - d = MIN(d, w - b->x - 1); - d = MIN(d, h - b->y - 1); + d = MIN(d, data->cols - b->x - 1); + d = MIN(d, data->rows - b->y - 1); if (d >= w_border) w_sum += 128; else - w_sum += 128 + (d - w_border) * w_mul / 256; + w_sum += 128 + (int)(w_border - d) * w_mul / 256; } w_total += w_sum; r->w_sum = w_sum; - x_avg /= r->count; - y_avg /= r->count; - DBG(" centroid=(%u %u)", (uns)x_avg, (uns)y_avg); + uns x_avg = x_sum / r->count; + uns y_avg = y_sum / r->count; + DBG(" centroid=(%u %u)", x_avg, y_avg); /* Compute normalized inertia */ u64 sum1 = 0, sum2 = 0, sum3 = 0; for (struct image_sig_block *b = r->blocks; b; b = b->next) { uns inc2 = isqr(x_avg - b->x) + isqr(y_avg - b->y); - uns inc1 = sqrt(inc2); + uns inc1 = fast_sqrt_u32(inc2); sum1 += inc1; sum2 += inc2; sum3 += inc1 * inc2; } - sig->reg[i].h[0] = CLAMP(image_sig_inertia_scale[0] * sum1 * ((3 * M_PI * M_PI) / 2) * pow(r->count, -1.5), 0, 65535); - sig->reg[i].h[1] = CLAMP(image_sig_inertia_scale[1] * sum2 * ((4 * M_PI * M_PI * M_PI) / 2) / ((u64)r->count * r->count), 0, 65535); - sig->reg[i].h[2] = CLAMP(image_sig_inertia_scale[2] * sum3 * ((5 * M_PI * M_PI * M_PI * M_PI) / 2) * pow(r->count, -2.5), 0, 65535); - + sig->reg[i].h[0] = CLAMP(image_sig_inertia_scale[0] * sum1 * ((3 * M_PI * M_PI) / 2) * pow(r->count, -1.5), 0, 255); + sig->reg[i].h[1] = CLAMP(image_sig_inertia_scale[1] * sum2 * ((4 * M_PI * M_PI * M_PI) / 2) / ((u64)r->count * r->count), 0, 255); + sig->reg[i].h[2] = CLAMP(image_sig_inertia_scale[2] * sum3 * ((5 * M_PI * M_PI * M_PI * M_PI) / 2) * pow(r->count, -2.5), 0, 255); + sig->reg[i].h[3] = (uns)x_avg * 127 / data->cols; + sig->reg[i].h[4] = (uns)y_avg * 127 / data->rows; } /* Compute average differences */ @@ -258,16 +261,16 @@ compute_image_signature(struct image_thread *thread UNUSED, struct image_signatu { uns d = 0; for (uns k = 0; k < IMAGE_REG_F; k++) - d += isqr(sig->reg[i].f[k] - sig->reg[j].f[k]); - df += sqrt(d); + d += image_sig_cmp_features_weights[k] * isqr(sig->reg[i].f[k] - sig->reg[j].f[k]); + df += fast_sqrt_u32(d); d = 0; for (uns k = 0; k < IMAGE_REG_H; k++) - d += isqr(sig->reg[i].h[k] - sig->reg[j].h[k]); - dh += sqrt(d); + d += image_sig_cmp_features_weights[k + IMAGE_REG_F] * isqr(sig->reg[i].h[k] - sig->reg[j].h[k]); + dh += fast_sqrt_u32(d); cnt++; } - sig->df = CLAMP(df / cnt, 1, 255); - sig->dh = CLAMP(dh / cnt, 1, 65535); + sig->df = CLAMP(df / cnt, 1, 0xffff); + sig->dh = CLAMP(dh / cnt, 1, 0xffff); } DBG("Average regions difs: df=%u dh=%u", sig->df, sig->dh); @@ -275,13 +278,17 @@ compute_image_signature(struct image_thread *thread UNUSED, struct image_signatu uns wa = 128, wb = 128; for (uns i = sig->len; --i > 0; ) { - struct image_sig_region *r = regions + i; - wa -= sig->reg[i].wa = CLAMP(r->count * 128 / blocks_count, 1, (int)(wa - i)); - wb -= sig->reg[i].wb = CLAMP(r->w_sum * 128 / w_total, 1, (int)(wa - i)); + struct image_sig_region *r = data->regions + i; + wa -= sig->reg[i].wa = CLAMP(r->count * 128 / data->blocks_count, 1, (int)(wa - i)); + wb -= sig->reg[i].wb = CLAMP(r->w_sum * 128 / w_total, 1, (int)(wb - i)); } sig->reg[0].wa = wa; sig->reg[0].wb = wb; + /* Store image dimensions */ + sig->cols = data->image->cols; + sig->rows = data->image->rows; + /* Dump regions features */ #ifdef LOCAL_DEBUG for (uns i = 0; i < sig->len; i++) @@ -291,9 +298,27 @@ compute_image_signature(struct image_thread *thread UNUSED, struct image_signatu DBG("region %u: features=%s", i, buf); } #endif +} - xfree(blocks); +void +image_sig_cleanup(struct image_sig_data *data) +{ + xfree(data->blocks); +} +int +compute_image_signature(struct image_context *ctx, struct image_signature *sig, struct image *image) +{ + struct image_sig_data data; + if (!image_sig_init(ctx, &data, image)) + return 0; + image_sig_preprocess(&data); + if (data.valid) + { + image_sig_segmentation(&data); + image_sig_detect_textured(&data); + } + image_sig_finish(&data, sig); + image_sig_cleanup(&data); return 1; } -