# rogerxavier-ocr-with-fastapi.hf.space import os import random#用于随机挑选封面 ##这个模型目前只适合确定文本框顺序后再识别,因为如果后面的 ##完整图片处理的反例 现在处理的图片是10\0.jpg # [[[953, 743], [987, 743], [987, 867], [953, 867]], [[917, 745], [951, 745], [951, 867], [917, 867]], [[881, 741], [918, 742], [915, 898], [877, 897]], [[843, 743], [879, 743], [879, 809], [843, 809]], [[629, 1058], [669, 1058], [669, 1210], [629, 1210]], [[549, 1227], [583, 1227], [583, 1381], [549, 1381]], [[535, 115], [563, 115], [563, 145], [535, 145]], [[535, 147], [563, 147], [563, 213], [535, 213]], [[507, 443], [539, 443], [539, 579], [507, 579]], [[505, 115], [533, 115], [533, 197], [505, 197]], [[511, 1225], [547, 1225], [547, 1321], [511, 1321]], [[475, 117], [503, 117], [503, 265], [475, 265]], [[467, 421], [503, 421], [503, 575], [467, 575]], [[419, 235], [447, 235], [447, 337], [419, 337]], [[387, 236], [417, 237], [414, 339], [385, 338]], [[209, 796], [242, 797], [239, 921], [206, 920]], [[175, 173], [205, 173], [205, 225], [175, 225]], [[177, 231], [205, 231], [205, 285], [177, 285]], [[103, 1153], [129, 1153], [129, 1223], [103, 1223]], [[41, 100], [108, 101], [104, 549], [36, 548]]] # ['就算是你', '没有圣剑', '也不可能有', '胜算', '就算如此', '我也不觉得', '做', ':做个', '·就不觉得', '老好人', '你可怕', '也要有个限度', '我很恐怖吗', '该说真是', '无药可救', '说的是呢', '这个', '但是', '为何?', '第二话让人怜爱'] import requests import tempfile import time import re #正则对话剔除非中文,保留'\n' from moviepy.audio.AudioClip import AudioArrayClip from moviepy.editor import * import cv2 import numpy as np import io import base64 import json from io import BytesIO import pandas as pd from PIL import Image import os from mutagen.mp3 import MP3 #读取音频获取时长 azure_speech_key = os.getenv('azure_speech_key') azure_service_region = os.getenv('azure_service_region') my_openai_key = os.getenv('my_openai_key') speech_synthesis_voice_name = "zh-CN-YunhaoNeural" ##云皓 print("azure key是",azure_speech_key) print("azure_service_region是",azure_service_region) print("my_openai_key",my_openai_key) from faker import Faker fake = Faker() fake_user_agent = fake.user_agent() fake_headers = { 'scheme': 'https', 'Accept': '*/*', 'Accept-Encoding': 'gzip, deflate, br, zstd', 'Accept-Language': 'zh-CN,zh;q=0.9', 'Dnt': '1', 'Priority': 'u=1, i', 'Sec-Ch-Ua': '"Chromium";v="124", "Google Chrome";v="124", "Not-A.Brand";v="99"', 'Sec-Ch-Ua-Mobile': '?0', 'Sec-Ch-Ua-Platform': '"Windows"', 'Sec-Fetch-Dest': 'empty', 'Sec-Fetch-Mode': 'cors', 'Sec-Fetch-Site': 'same-origin', 'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36' } fake_headers['User-Agent'] = fake_user_agent #通过去水印完整漫画图片->获取相应的对话框图片->获取对话框文字->返回对话框文字 def get_image_copywrite(image_path:"图片路径(包含后缀)",dialog_cut_path:"对话框切割路径")->"返回漫画关联对话框识别后得到的文案str(原文即可),也可能是none": def extract_chinese(text:str)->str: #剔除除了 '\n'外的非中文字符 chinese_pattern = re.compile("[\u4e00-\u9fa5]+") # 匹配中文字符的正则表达式 chinese_text = "" for char in text: if char == '\n' or re.match(chinese_pattern, char): chinese_text += char return chinese_text dialog_texts = '' associate_dialog_img = get_associate_dialog(image_path=image_path,dialog_cut_path=dialog_cut_path) if len(associate_dialog_img)!=0: #如果有对应的对话框 for dialog_img_path in associate_dialog_img: cur_dialog_texts = get_sorted_dialog_text(image_path=dialog_img_path,fake_headers = fake_headers)#一个对话框的文字list if cur_dialog_texts is not None: for dialog_text in cur_dialog_texts: # dialog_texts += dialog_text dialog_texts += extract_chinese(dialog_text) #因为已经在数组中加入了\n 换行,这里就不用加了 else: print(dialog_img_path+"识别是空-可能是有问题") return dialog_texts return None#不规范图片不请求,直接返回none #通过传入无水印漫画图片对话框路径,得到关联的对话框图片list def get_associate_dialog(image_path:"图片路径(包含后缀)",dialog_cut_path:"对话框切割路径")->"返回漫画关联对话框list,也可能是空的list": image_name = os.path.splitext(os.path.basename(image_path))[0] image_name_format = '{:03d}'.format(int(image_name)) associated_dialogs = [] for root, _, files in os.walk(dialog_cut_path): for file in files: if file.startswith(image_name_format) and file.endswith('.jpg'): associated_dialogs.append(os.path.join(root, file)) return associated_dialogs def merge_sublists(lists): merged = [] for sublist in lists: found = False for m in merged: if any(elem in sublist for elem in m): m.extend(elem for elem in sublist if elem not in m) found = True break if not found: merged.append(sublist) return merged # 任意两框进行中心高度差和中心宽度差比较,如果xy都相近,那么认为是同一个框的对话,加入一个对话数组里面, # 最终将漫画块分成几个对话框数组,然后再对数组间进行从上到下,从右到左排序 # 定义一个函数来寻找相关的点并加入新的list def find_associate_text(sorted_indices,centers,sorted_coordinates,boxInfo): associate_text_list = [] related_groups = [] for i in range(len(sorted_indices) - 1): for j in range(i+1 , len(sorted_indices)): if (abs(centers[sorted_indices[i]][1] - centers[sorted_indices[j]][1]) < abs( (sorted_coordinates[i][2][1] - sorted_coordinates[i][0][1])) / 3) \ and (abs(centers[sorted_indices[i]][0] - centers[sorted_indices[j]][0]) < abs( (sorted_coordinates[i][2][0] - sorted_coordinates[i][0][0])) * 1.5): # Check if the points i and j are already in the same related group found = False for group in related_groups: if i in group or j in group: group.add(i) group.add(j) found = True break if not found: related_groups.append({i, j}) for group in related_groups: text_group = [] for idx in group: text_group.append(boxInfo['Text'][str(sorted_indices[idx])])#这里加入的是排序后的索引 associate_text_list.append(text_group) return merge_sublists(associate_text_list),related_groups #先对组内对话从右到左排序,处理反馈到related_groups (因为sorted_indices本身就是从右到左,从上到下排序后的) # 这个记录的顺序改变,最后sorted_text = [boxInfo['Text'][str(i)] for i in sorted_indices]就可以得到正确的顺序 #要保证一个List中的组内有序和组间有序，通常应该先排序组内，然后再保持组间有序 def sort_associate_text_list(sorted_indices:list,related_groups:list,boxCoordinates,centers)->list: sorted_groups = [] # 返回组内排序后的 sorted_groups for group in related_groups: group = list(group) # 将集合转换为列表 isVertical = False isCross = False # 前提是竖框->使用 lambda 函数按照中心点坐标的 x 值对 group 中的元素进行排序，使得x大的(靠右的)在前面 for idx in group: if (boxCoordinates[sorted_indices[idx]][2][0] - boxCoordinates[sorted_indices[idx]][0][0]) > ( boxCoordinates[sorted_indices[idx]][2][1] - boxCoordinates[sorted_indices[idx]][0][1]): # 这里是宽>高,说明是横框 isCross =True pass # 你可以在这里添加你想要执行的代码 else: # 这里宽<高,说明是竖框 isVertical = True pass # 你可以在这里添加你想要执行的代码 if isVertical: group.sort(key=lambda idx: centers[sorted_indices[idx]][0], reverse=True) if isCross: group.sort(key=lambda idx: centers[sorted_indices[idx]][1], reverse=False) sorted_groups.append(group) return sorted_groups #再对组间对话先上后下,从右到左排序,同时将单独对话加入合适位置.返回排序后的related_groups def sort_dialog_list(sorted_indices:list,related_groups:list,sorted_coordinates)->list: sorted_groups = [] related_groups_copy = related_groups.copy() sorted_indices_copy = sorted_indices.copy() added = {} # 返回组内排序后的 sorted_groups # 任意两框进行加权高度差值比较,然后交换顺序,而不是只遍历一遍交换,如果y中心点差在1/3 文本框长度下认为相同,这时按照x从右往左顺序看 for i in range(len(sorted_indices) - 1): if (( sorted_coordinates[i][2][0] - sorted_coordinates[i][0][0]) < ( sorted_coordinates[i][2][1] - sorted_coordinates[i][0][1])): # 竖框情况下(宽小于高),依次加入元素,加到在组中的那么后序按组顺序加,然后继续（再碰到不加） pass #竖框不动,横框剔除,后序不在次循环中储粮 else: sorted_indices_copy.remove(i) # 横框情况下(宽大于高)#横框干脆不读了(从sorted_indices_copy中剔除),太影响了 for idx in sorted_indices_copy: added[idx] = False for group in related_groups_copy: if idx in group: sorted_groups.append(group) related_groups_copy.remove(group) added[idx] = True break if not added[idx]: sorted_groups.append(idx) # 创建一个新列表来存储不应该单独存在的元素,并且游离的元素也变[]包裹 filtered_data = [] data = sorted_groups for item in data: if isinstance(item, list): # 如果元素是列表，则将其添加到新列表中 filtered_data.append(item) else: # 如果元素不是列表，则检查是否存在于其他子项数组中，如果不存在则添加到新列表中 is_in_sublist = False for sublist in data: if isinstance(sublist, list) and item in sublist: is_in_sublist = True break if not is_in_sublist: filtered_data.append([item]) return filtered_data def get_sorted_dialog_text(image_path:"包含后缀的文件路径",fake_headers:dict=None)->"返回排序后的text list(一列或者几列话,反正是一个框的内容,几句不清楚,一个框的list当一次文案就行) 或者失败请求返回none": # 创建一个Session对象 session = requests.Session() if fake_headers is not None: session.headers = fake_headers image_bytes = open(image_path, 'rb') files = { "image": image_bytes, } try: resp = session.post("https://rogerxavier-ocr-with-fastapi.hf.space/getCoordinates", files=files)#用faker header看看会不会还限制 #先json转换，0为坐标list合集,1为 boxid和text合集 boxCoordinates , boxInfo = resp.json()[0],resp.json()[1] #分别是list和dict类型 print("ofa ocr识别漫画块成功返回") print("boxCoordinates是",boxCoordinates) print("boxInfo是",boxInfo) # 计算文本框的中心点，以便按照从右往左，从上往下的顺序进行排序 centers = [((box[0][0] + box[2][0]) / 2, (box[0][1] + box[2][1]) / 2) for box in boxCoordinates] # 根据每个元组的第一个元素进行降序排序，如果第一个元素相同时，则根据第二个元素进行升序排序。即先关注y后关注x(更重上下) sorted_indices = sorted(range(len(centers)), key=lambda i: ( centers[i][1],-centers[i][0])) # # 即先关注x后关注y(更注重从右到左) # sorted_indices = sorted(range(len(centers)), key=lambda i: ( -centers[i][0],centers[i][1])) # 获取排序后的文本框坐标和对应的文字 sorted_coordinates = [boxCoordinates[i] for i in sorted_indices] # 调用函数并打印结果 associate_text_list,related_groups = find_associate_text(sorted_indices,centers,sorted_coordinates,boxInfo) related_groups = sort_associate_text_list(sorted_indices,related_groups,boxCoordinates,centers) #print("组内排序后的related_groups是",related_groups) #[[3, 4], [7, 5, 6], [10, 9], [11, 12, 13], [15, 16, 14]] related_groups_in_sorted_indices = [] for group in related_groups: related_groups_in_sorted_indices_item = [] for idx in group: related_groups_in_sorted_indices_item.append(sorted_indices[idx])# 这里加入的是排序后的索引 related_groups_in_sorted_indices.append(related_groups_in_sorted_indices_item) #print("related_groups_in_sorted_indices是",related_groups_in_sorted_indices) #related_groups_in_sorted_indices->[[7, 6], [3, 2, 4], [9, 10], [11, 13, 12], [15, 16, 14]]-> #期望结果[[0],[3, 2, 4],[1],[5],[7, 6],[8], [9, 10],[11, 13, 12], [15, 16, 14]] related_groups = sort_dialog_list(sorted_indices,related_groups,sorted_coordinates) #print("related_groups组件排序后是:",related_groups) # 将子列表中的数字提取出来组成一个新的列表(纯数字),去除子项间的[]， # 如[[3, 4], [7, 5, 6], [10, 9], [11, 12, 13], [15, 16, 14]] ->[3, 4, 7, 5, 6, 10, 9, 11, 12, 13, 15, 16, 14] flattened_list = [num for sublist in related_groups for num in sublist] added_indices = set() sorted_text = [] for i in flattened_list: for sublist in related_groups: if i in sublist: if i == sublist[-1] and i not in added_indices: sorted_text.append(boxInfo['Text'][str(sorted_indices[i])] + '\n') added_indices.add(i) elif i not in added_indices: sorted_text.append(boxInfo['Text'][str(sorted_indices[i])]) added_indices.add(i) #print("不完整的sorted_text是",sorted_text) # 不用在最后一个项末尾添加"\n",从而隔开其他的漫画块对话(因为总会有最后一个子块,因而上述方式就可以加上了) sorted_coordinates = [boxCoordinates[i] for i in sorted_indices] print(sorted_coordinates) print(sorted_text) return sorted_text except Exception as e: print("ofa ocr图片请求出现问题") print(e) return None #通过文字获取音频 def get_audio_data(text:str)-> "返回audio data io句柄, duration(也有可能包含无效字符导致生成音频400错误)": # Creates an instance of a speech config with specified subscription key and service region. speech_key = azure_speech_key service_region = azure_service_region voiceText = text url = f"https://{service_region}.tts.speech.microsoft.com/cognitiveservices/v1" headers = { "Ocp-Apim-Subscription-Key": speech_key, "Content-Type": "application/ssml+xml", "X-Microsoft-OutputFormat": "audio-16khz-128kbitrate-mono-mp3", "User-Agent": "curl" } ssml_text = ''' {voiceText} '''.format(voiceName=speech_synthesis_voice_name,voiceText = voiceText) response = requests.post(url, headers=headers, data=ssml_text.encode('utf-8')) if response.status_code == 200: # 创建临时文件 -当前路径下面 try: with tempfile.NamedTemporaryFile(dir='/mp3_out/',delete=False) as temp_file: temp_file.write(response.content) temp_file.close() audio = MP3(temp_file.name) # 获取音频时长（单位为秒） audio_duration_seconds = audio.info.length #int即可 # 在这里完成您对文件的操作，比如返回文件名 file_name = temp_file.name return file_name, audio_duration_seconds except Exception as e: print("可能遇到mp3 can not sync to MPEG frame错误,总之音频能获取到但是不能识别",e) return None,None#这种也返回none告知错误不要管了 else: print("Error: Failed to synthesize audio. Status code:", response.status_code) return None,None # 补零函数，将数字部分补齐为指定长度 def zero_pad(s, length): return s.zfill(length) def gpt_polish(text:str)->"通过gpt润色str文案并返回str新文案,或者gpt请求失败none": # Set your OpenAI API key api_key = my_openai_key # Define the headers headers = { 'Authorization': f'Bearer {api_key}', 'Content-Type': 'application/json', } # Chat Completions request data data = { 'model': 'gpt-3.5-turbo', # Replace with your chosen model 'messages': [ {'role': 'system', 'content': "你是一个assistant,能够根据user发送的漫画中提取的对话文字,生成一个短视频中一帧的文案(1-2句话)"}, {'role': 'user', 'content': text} ] } try: response = requests.post('https://api.yingwu.lol/v1/chat/completions', headers=headers, data=json.dumps(data)) print("gpt请求的结果是",response.text) print("润色后文案是:"+response.json()['choices'][0]['message']['content']) return response.json()['choices'][0]['message']['content'] except Exception as e: print("gpt润色文案失败:") print(e) return None if __name__ == '__main__': # 获取存放去水印漫画图片的路径 ---放这里是因为获取对话文字时需要和原图关联 img_path = 'manga1' # 获取切割后的文本框路径 dialog_img_path = 'manga12' #获取漫画原图无水印的加入image_files，并排序 subdir_path = os.path.join(os.getcwd(), img_path) # 对话图片经过加入list并补0确定顺序 image_files = [] for root, dirs, files in os.walk(subdir_path): for file in files: if file.endswith(".jpg") or file.endswith(".png"): image_files.append(os.path.relpath(os.path.join(root, file))) # 对对话框文件名中的数字部分进行补零操作-这样顺序会正常 image_files.sort( key=lambda x: zero_pad(''.join(filter(str.isdigit, os.path.splitext(os.path.basename(x))[0])), 3)) dialog_subdir_path = os.path.join(os.getcwd(), dialog_img_path) # 对话图片经过加入list并补0确定顺序 dialog_image_files = [] for root, dirs, files in os.walk(dialog_subdir_path): for file in files: if file.endswith(".jpg") or file.endswith(".png"): dialog_image_files.append(os.path.relpath(os.path.join(root, file))) # 对对话框文件名中的数字部分进行补零操作-这样顺序会正常 dialog_image_files.sort( key=lambda x: zero_pad(''.join(filter(str.isdigit, os.path.splitext(os.path.basename(x))[0])), 3)) # 对话图片经过加入list并补0确定顺序 ###音视频相关参数------------------------------------------------------------------------------------- ##这个是临时生成音频文件的全局变量--方便后续删除 filename = '' # 视频分辨率和帧率 # 获取第一张图片的尺寸 #image = Image.open(image_files[0]) #随机一个作为cover image=Image.open(random.choice(image_files)) width, height = image.size #使用图片的size作为宽高 #读取第一个图片作为cover保存到cover/0.jpg # 定义要保存的文件路径 save_path = os.path.join("cover", "0.jpg") # 保存图片文件 image.save(save_path) #读取第一个图片作为cover保存到cover/0.jpg fps = 30 font_path = '1.ttf' # 设置字体以防默认字体无法同时处理中英文 # 创建视频编辑器 video_clips = [] ###音视频相关参数------------------------------------------------------------------------------------- #因为是根据原图无水印的进行遍历,所以处理前要进行筛选，只处理能找到相应对话框图片的原图 filtered_image_files = [] for image_path in image_files: dialog_list = get_associate_dialog(image_path, dialog_img_path) if dialog_list: filtered_image_files.append(image_path) image_files = filtered_image_files for idx, image_file in enumerate(image_files): print("现在处理的图片是"+image_file) #后面是视音频生成部分-这里图片需要用到完整的去水印的而不是对话框用于识别的 img = cv2.imread(image_file) img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) ##只支持英文路径 ##获取当前图片对应的对话框识别文字(还需gpt处理后作为字幕文案) cur_copywrite = get_image_copywrite(image_file,dialog_img_path) # image_file就是6.jpg了 #cur_copywrite = gpt_polish(cur_copywrite)#不用gpt，只用新版漫画块得到的100%识别原文即可 if cur_copywrite is not None: ##获取当前图片对应的临时音频文件名称和文案时长 # filename, duration = get_audio_data(cur_copywrite) filename, duration = get_audio_data(cur_copywrite)#这里是一个原图的全部文案,不是一个漫画块的,不能在这里加\n断开不同漫画块的对话 if filename is not None: print("存放临时mp3文件的路径是",filename) #不含字幕版 clip = ImageClip(img).set_duration(duration).resize((width, height)) # 去掉添加字幕的部分(原文太长了,再加上音频都是一整个原图(即多个漫画块)的全部内容,也没法分割) audio_clip = AudioFileClip(filename) clip = clip.set_audio(audio_clip) video_clips.append(clip) #不含字幕版 else: pass ##音频特殊字符或者其他原因无法生成跳过 video = concatenate_videoclips(video_clips) # 保存视频 video.write_videofile('mp4_out/output_video.mp4', fps=fps,temp_audiofile="mp3_out/temp.mp3") # # 在文件关闭后删除临时文件 print("删除临时mp3文件", filename) os.remove(filename)