picpocket / rvm_processor.py
chawin.chen
init
cd5aabe
import os
import cv2
import numpy as np
import torch
from torchvision import transforms
import config
from config import logger
class RVMProcessor:
"""RVM (Robust Video Matting) 抠图处理器"""
def __init__(self):
self.model = None
self.available = False
self.device = "cpu" # 默认使用CPU,如果有GPU可以设置为"cuda"
try:
# 仅从本地加载,不使用网络
local_repo = getattr(config, 'RVM_LOCAL_REPO', '')
weights_path = getattr(config, 'RVM_WEIGHTS_PATH', '')
if not local_repo or not os.path.isdir(local_repo):
raise RuntimeError("RVM_LOCAL_REPO not set or invalid. Please set env RVM_LOCAL_REPO to local RobustVideoMatting repo path (with hubconf.py)")
if not weights_path or not os.path.isfile(weights_path):
raise RuntimeError("RVM_WEIGHTS_PATH not set or file not found. Please set env RVM_WEIGHTS_PATH to local RVM weights file path")
logger.info(f"Loading RVM model {config.RVM_MODEL} from local repo: {local_repo}")
# 使用本地仓库构建模型,禁用预训练以避免联网
self.model = torch.hub.load(local_repo, config.RVM_MODEL, source='local', pretrained=False)
# 加载本地权重
state = torch.load(weights_path, map_location=self.device)
if isinstance(state, dict) and 'state_dict' in state:
state = state['state_dict']
missing, unexpected = self.model.load_state_dict(state, strict=False)
# 迁移到设备并设置评估模式
self.model = self.model.to(self.device).eval()
self.available = True
logger.info("RVM background removal processor initialized successfully (local mode)")
if missing:
logger.warning(f"RVM weights missing keys: {list(missing)[:5]}... total={len(missing)}")
if unexpected:
logger.warning(f"RVM weights unexpected keys: {list(unexpected)[:5]}... total={len(unexpected)}")
except Exception as e:
logger.error(f"RVM background removal processor initialization failed: {e}")
self.available = False
def is_available(self) -> bool:
"""检查RVM处理器是否可用"""
return self.available and self.model is not None
def remove_background(self, image: np.ndarray, background_color: tuple = None) -> np.ndarray:
"""
使用RVM移除图片背景
:param image: 输入的OpenCV图像(BGR格式)
:param background_color: 替换的背景颜色(BGR格式),如果为None则保持透明背景
:return: 处理后的图像
"""
if not self.is_available():
raise Exception("RVM抠图处理器不可用")
try:
logger.info("Starting to remove background using RVM...")
# 保存原始图像尺寸
original_height, original_width = image.shape[:2]
# 将OpenCV图像(BGR)转换为RGB格式
image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
# 转换为tensor
src = transforms.ToTensor()(image_rgb).unsqueeze(0).to(self.device)
# 推理
rec = [None] * 4
with torch.no_grad():
fgr, pha, *rec = self.model(src, *rec, downsample_ratio=0.25)
# 转换为numpy数组
fgr = (fgr[0].permute(1, 2, 0).cpu().numpy() * 255).astype(np.uint8) # (H,W,3)
pha = (pha[0, 0].cpu().numpy() * 255).astype(np.uint8) # (H,W)
# 检查尺寸是否匹配,如果不匹配则调整
if fgr.shape[:2] != (original_height, original_width):
fgr = cv2.resize(fgr, (original_width, original_height))
pha = cv2.resize(pha, (original_width, original_height))
if background_color is not None:
# 如果指定了背景颜色,创建纯色背景
# 将前景图像转换为BGR格式
fgr_bgr = cv2.cvtColor(fgr, cv2.COLOR_RGB2BGR)
# 创建背景图像
background = np.full((original_height, original_width, 3), background_color, dtype=np.uint8)
# 使用alpha混合
alpha = pha.astype(np.float32) / 255.0
alpha = np.stack([alpha] * 3, axis=-1)
result = (fgr_bgr * alpha + background * (1 - alpha)).astype(np.uint8)
else:
# 保持透明背景,转换为BGRA格式
fgr_bgr = cv2.cvtColor(fgr, cv2.COLOR_RGB2BGR)
rgba = np.dstack((fgr_bgr, pha)) # (H,W,4)
result = rgba
logger.info("RVM background removal completed")
return result
except Exception as e:
logger.error(f"RVM background removal failed: {e}")
raise Exception(f"背景移除失败: {str(e)}")
def create_id_photo(self, image: np.ndarray, background_color: tuple = (255, 255, 255)) -> np.ndarray:
"""
创建证件照(移除背景并添加纯色背景)
:param image: 输入的OpenCV图像
:param background_color: 背景颜色,默认白色(BGR格式)
:return: 处理后的证件照
"""
logger.info(f"Starting to create ID photo, background color: {background_color}")
# 移除背景并添加指定颜色背景
id_photo = self.remove_background(image, background_color)
logger.info("ID photo creation completed")
return id_photo