#!/usr/bin/env python3 # -*- coding: utf-8 -*- """ 分析古籍OCR JSON数据 - 基于间隔检测的物理列聚合 """ import json import numpy as np from collections import defaultdict def analyze_physical_columns(json_path): with open(json_path, 'r', encoding='utf-8') as f: data = json.load(f) print("=" * 80) print(f"文件: {data['FileName']}") print(f"尺寸: {data['Width']} × {data['Height']}") print(f"版式: 10列 × 25行") print("=" * 80) chars = data['chars'] char_marking = data['charMarking'] coors = data['coors'] line_ids = data['line_ids'] # 按line_id分组 logical_columns = defaultdict(list) for i in range(len(chars)): logical_columns[line_ids[i]].append({ 'index': i, 'char': chars[i], 'x1': coors[i][0], 'y1': coors[i][1], 'x2': coors[i][2], 'y2': coors[i][3], 'x_center': (coors[i][0] + coors[i][2]) / 2, 'y_center': (coors[i][1] + coors[i][3]) / 2, 'is_small': len(char_marking[i]) > 0 }) # 计算每个逻辑列的信息 logical_col_info = [] for line_id in sorted(logical_columns.keys()): col_chars = logical_columns[line_id] x_centers = [c['x_center'] for c in col_chars] x_avg = np.mean(x_centers) char_count = len(col_chars) small_count = sum(1 for c in col_chars if c['is_small']) logical_col_info.append({ 'line_id': line_id, 'x_avg': x_avg, 'chars': col_chars, 'char_count': char_count, 'small_count': small_count, 'is_all_small': small_count == char_count }) # 按x坐标排序（从右到左） logical_col_info.sort(key=lambda c: c['x_avg'], reverse=True) print("\n逻辑列x坐标（从右到左）:") for lc in logical_col_info: small_mark = "(小字)" if lc['is_all_small'] else "" print(f" line_id={lc['line_id']}: x={lc['x_avg']:.0f} {small_mark}") # 基于间隔检测来聚合物理列 print("\n" + "=" * 80) print("基于间隔的物理列聚合:") print("-" * 80) # 新策略： # 1. 所有逻辑列一起考虑（大字和小字） # 2. 如果相邻逻辑列的x坐标差距小于阈值（如150px），它们属于同一物理列 # 3. 如果差距大于阈值，就是新的物理列 # 计算相邻逻辑列的间隔 print("所有逻辑列间隔分析:") gaps = [] for i in range(len(logical_col_info) - 1): gap = logical_col_info[i]['x_avg'] - logical_col_info[i+1]['x_avg'] gaps.append(gap) is_small_1 = "(小)" if logical_col_info[i]['is_all_small'] else "" is_small_2 = "(小)" if logical_col_info[i+1]['is_all_small'] else "" print(f" line_id={logical_col_info[i]['line_id']}{is_small_1} -> line_id={logical_col_info[i+1]['line_id']}{is_small_2}: {gap:.0f}px") # 分析间隔分布 print(f"\n间隔统计: min={min(gaps):.0f}, max={max(gaps):.0f}, avg={np.mean(gaps):.0f}, std={np.std(gaps):.0f}") # 使用阈值：小于150px认为是同一物理列（双行小字的左右列差约112px） # 大于150px认为是不同物理列 MERGE_THRESHOLD = 150 # 小于这个值就合并 physical_columns = [] current_group = [logical_col_info[0]] for i in range(1, len(logical_col_info)): gap = current_group[-1]['x_avg'] - logical_col_info[i]['x_avg'] if gap < MERGE_THRESHOLD: # 间隔小，属于同一物理列 current_group.append(logical_col_info[i]) else: # 间隔大，新的物理列 physical_columns.append(current_group) current_group = [logical_col_info[i]] physical_columns.append(current_group) print(f"\n聚合结果: {len(physical_columns)} 个物理列") print("-" * 80) # 计算物理列的实际中心位置 pc_centers = [] for pc in physical_columns: x_values = [lc['x_avg'] for lc in pc] center = np.mean(x_values) pc_centers.append(center) # 计算列间距 print("\n物理列间距:") for i in range(len(pc_centers) - 1): gap = pc_centers[i] - pc_centers[i+1] print(f" 物理列{i+1} -> 物理列{i+2}: {gap:.0f}px") avg_gap = np.mean([pc_centers[i] - pc_centers[i+1] for i in range(len(pc_centers)-1)]) print(f" 平均间距: {avg_gap:.0f}px") # 推断在10列网格中的位置 canvas_width = data['Width'] cell_width = canvas_width / 10 print(f"\n在10列网格中的映射 (列宽={cell_width:.0f}px):") for pi, (pc, center) in enumerate(zip(physical_columns, pc_centers)): grid_col = round((canvas_width - center) / cell_width) if grid_col < 1: grid_col = 1 if grid_col > 10: grid_col = 10 print(f" 物理列{pi+1} (中心x={center:.0f}) -> 网格第{grid_col}列") # 显示每个物理列的详细信息 print("\n" + "=" * 80) print("物理列详情:") print("-" * 80) cell_height = data['Height'] / 25 for pi, pc in enumerate(physical_columns): line_ids_in_pc = [lc['line_id'] for lc in pc] total_chars = sum(lc['char_count'] for lc in pc) total_small = sum(lc['small_count'] for lc in pc) print(f"\n物理列 {pi+1}:") print(f" 包含line_ids: {line_ids_in_pc}") print(f" 总字数: {total_chars} (小字: {total_small})") # 合并所有字符，按y排序 all_chars = [] for lc in pc: all_chars.extend(lc['chars']) all_chars.sort(key=lambda c: c['y_center']) # 分析行分配 first_y = all_chars[0]['y_center'] start_row = int(first_y / cell_height) print(f" 起始行: 第{start_row+1}行 (y={first_y:.0f})") # 显示内容结构 big_chars_content = ''.join([c['char'] for c in all_chars if not c['is_small']]) small_chars_content = ''.join([c['char'] for c in all_chars if c['is_small']]) print(f" 大字: {big_chars_content[:20]}{'...' if len(big_chars_content) > 20 else ''}") if small_chars_content: print(f" 小字: {small_chars_content}") # 如果有小字，分析配对 small_chars = [c for c in all_chars if c['is_small']] if small_chars: x_centers = [c['x_center'] for c in small_chars] x_threshold = np.mean(x_centers) right_chars = sorted([c for c in small_chars if c['x_center'] >= x_threshold], key=lambda c: c['y_center']) left_chars = sorted([c for c in small_chars if c['x_center'] < x_threshold], key=lambda c: c['y_center']) print(f" 双行小字配对:") print(f" 右列({len(right_chars)}字): {''.join([c['char'] for c in right_chars])}") print(f" 左列({len(left_chars)}字): {''.join([c['char'] for c in left_chars])}") if __name__ == '__main__': analyze_physical_columns('/home/yuuko/test/0011B.json')