// Copyright 2011 Google Inc. All Rights Reserved. // // Use of this source code is governed by a BSD-style license // that can be found in the COPYING file in the root of the source // tree. An additional intellectual property rights grant can be found // in the file PATENTS. All contributing project authors may // be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // Quantize levels for specified number of quantization-levels ([2, 256]). // Min and max values are preserved (usual 0 and 255 for alpha plane). // // Author: Skal (pascal.massimino@gmail.com) #include <assert.h> #include "./quant_levels_utils.h" #define NUM_SYMBOLS 256 #define MAX_ITER 6 // Maximum number of convergence steps. #define ERROR_THRESHOLD 1e-4 // MSE stopping criterion. // ----------------------------------------------------------------------------- // Quantize levels. int QuantizeLevels(uint8_t* const data, int width, int height, int num_levels, uint64_t* const sse) { int freq[NUM_SYMBOLS] = { 0 }; int q_level[NUM_SYMBOLS] = { 0 }; double inv_q_level[NUM_SYMBOLS] = { 0 }; int min_s = 255, max_s = 0; const size_t data_size = height * width; int i, num_levels_in, iter; double last_err = 1.e38, err = 0.; const double err_threshold = ERROR_THRESHOLD * data_size; if (data == NULL) { return 0; } if (width <= 0 || height <= 0) { return 0; } if (num_levels < 2 || num_levels > 256) { return 0; } { size_t n; num_levels_in = 0; for (n = 0; n < data_size; ++n) { num_levels_in += (freq[data[n]] == 0); if (min_s > data[n]) min_s = data[n]; if (max_s < data[n]) max_s = data[n]; ++freq[data[n]]; } } if (num_levels_in <= num_levels) goto End; // nothing to do! // Start with uniformly spread centroids. for (i = 0; i < num_levels; ++i) { inv_q_level[i] = min_s + (double)(max_s - min_s) * i / (num_levels - 1); } // Fixed values. Won't be changed. q_level[min_s] = 0; q_level[max_s] = num_levels - 1; assert(inv_q_level[0] == min_s); assert(inv_q_level[num_levels - 1] == max_s); // k-Means iterations. for (iter = 0; iter < MAX_ITER; ++iter) { double q_sum[NUM_SYMBOLS] = { 0 }; double q_count[NUM_SYMBOLS] = { 0 }; int s, slot = 0; // Assign classes to representatives. for (s = min_s; s <= max_s; ++s) { // Keep track of the nearest neighbour 'slot' while (slot < num_levels - 1 && 2 * s > inv_q_level[slot] + inv_q_level[slot + 1]) { ++slot; } if (freq[s] > 0) { q_sum[slot] += s * freq[s]; q_count[slot] += freq[s]; } q_level[s] = slot; } // Assign new representatives to classes. if (num_levels > 2) { for (slot = 1; slot < num_levels - 1; ++slot) { const double count = q_count[slot]; if (count > 0.) { inv_q_level[slot] = q_sum[slot] / count; } } } // Compute convergence error. err = 0.; for (s = min_s; s <= max_s; ++s) { const double error = s - inv_q_level[q_level[s]]; err += freq[s] * error * error; } // Check for convergence: we stop as soon as the error is no // longer improving. if (last_err - err < err_threshold) break; last_err = err; } // Remap the alpha plane to quantized values. { // double->int rounding operation can be costly, so we do it // once for all before remapping. We also perform the data[] -> slot // mapping, while at it (avoid one indirection in the final loop). uint8_t map[NUM_SYMBOLS]; int s; size_t n; for (s = min_s; s <= max_s; ++s) { const int slot = q_level[s]; map[s] = (uint8_t)(inv_q_level[slot] + .5); } // Final pass. for (n = 0; n < data_size; ++n) { data[n] = map[data[n]]; } } End: // Store sum of squared error if needed. if (sse != NULL) *sse = (uint64_t)err; return 1; }