ASJPush/Object/NovaImageTagObject.py

# -*- encoding: utf-8 -*-
"""
@File    : NovaImageTagObject.py
@Time    : 2025/8/29 09:03
@Author  : stephen
@Email   : zhangdongming@asj6.wecom.work
@Software: PyCharm
"""
import base64
import io
import json
import logging
import re

import boto3
from PIL import Image

LOGGER = logging.getLogger('time')

# --- 配置信息 ---
MODEL_ID = "us.amazon.nova-lite-v1:0"


class NovaImageTagObject(object):
    def __init__(self, aws_access_key_id, secret_access_key, region_name):
        self.bedrock = boto3.client(
            'bedrock-runtime',
            aws_access_key_id=aws_access_key_id,
            aws_secret_access_key=secret_access_key,
            region_name=region_name
        )

    @staticmethod
    def safe_json_load(json_string):
        """
        一个更健壮的JSON解析函数，尝试修复常见的模型输出格式问题。
        """
        try:
            # 寻找被代码块包围的JSON
            json_match = re.search(r'```json\s*([\s\S]*?)\s*```', json_string)
            if json_match:
                json_string = json_match.group(1)

            # 寻找常规的JSON对象或数组
            json_match = re.search(r'\{.*\}|\[.*\]', json_string, re.DOTALL)
            if json_match:
                json_string = json_match.group(0)

            return json.loads(json_string)
        except json.JSONDecodeError:
            LOGGER.error("JSON解析失败，尝试修复...")
            try:
                json_string = re.sub(r"(\w+):", r'"\1":', json_string)
                json_string = json_string.replace("'", '"')
                return json.loads(json_string)
            except Exception as e:
                LOGGER.error(f"无法解析模型返回的JSON: {e}")
                return None
        except Exception as e:
            LOGGER.error(f"发生未知解析错误: {e}")
            return None

    @staticmethod
    def format_and_convert_detections(nova_detections: list) -> list:
        """
        将Nova模型返回的坐标转换为您指定的详细格式，包含原始坐标和Rekognition比例。
        """
        formatted_results = []
        if not isinstance(nova_detections, list):
            return []

        for item in nova_detections:
            if not isinstance(item, dict): continue

            label = list(item.keys())[0]
            nx1, ny1, nx2, ny2 = item[label]

            left = nx1 / 1000.0
            top = ny1 / 1000.0
            width = (nx2 - nx1) / 1000.0
            height = (ny2 - ny1) / 1000.0

            formatted_results.append({
                "x1": nx1, "x2": nx2, "y1": ny1, "y2": ny2,
                "Width": f"{width:.5f}", "Height": f"{height:.5f}",
                "Top": f"{top:.5f}", "Left": f"{left:.5f}",
                "class": label
            })
        return formatted_results

    @staticmethod
    def normalize_b64(b64_str: str) -> str:
        """清理并补齐base64字符串"""
        if not b64_str:
            return ""
        b64_str = re.sub(r"^data:image/[^;]+;base64,", "", b64_str)
        b64_str = b64_str.strip().replace("\n", "").replace(" ", "")
        # 补齐Base64填充
        padding = 4 - (len(b64_str) % 4)
        if padding and padding != 4:
            b64_str += "=" * padding
        return b64_str

    def process_image_batch(self, base64_images: list, categories: list, uid=''):
        if not base64_images:
            LOGGER.error(f"{uid}错误: 未提供图片数据。")
            return {}

        image_contents = []
        for idx, b64_image in enumerate(base64_images, start=1):
            try:
                # 规范化base64
                b64_image = self.normalize_b64(b64_image)
                if not b64_image:
                    raise ValueError("空的base64字符串")

                # 解码为二进制
                img_bytes = base64.b64decode(b64_image)

                # 使用PIL处理图像
                image = Image.open(io.BytesIO(img_bytes))

                # 转换为RGB模式（如果需要）
                if image.mode != 'RGB':
                    image = image.convert('RGB')

                # 转换为WebP格式
                buffer = io.BytesIO()
                image.save(buffer, format="webp", quality=90)
                img_bytes = buffer.getvalue()

                # 直接传递二进制数据给Bedrock
                image_contents.append({
                    "image": {"format": "webp", "source": {"bytes": img_bytes}}
                })

                LOGGER.info(f"{uid} 第{idx}张图处理成功, 格式=webp, 大小={len(img_bytes)}B")

            except Exception as e:
                LOGGER.error(f"{uid} 第{idx}张图处理失败，已跳过: {repr(e)}")
                image_contents.append(None)  # 添加占位符

        if not image_contents:
            LOGGER.error(f"{uid}错误: 所有图片均无法处理。")
            return {}

        category_str = ", ".join([f'"{cat.lower()}"' for cat in categories])
        num_images = len(image_contents)

        # --- 关键改动：为多图片设计的全新Prompt ---
        prompt = f"""
    You have been provided with {num_images} images. Analyze each image sequentially.
    For each image, detect bounding boxes of objects from the following categories: {category_str}.
    Your output MUST be a single, valid JSON object.
    The keys of this object should be "image_0", "image_1", ..., "image_{num_images - 1}", corresponding to the first, second, and subsequent images provided.
    The value for each key must be a list of detected objects for that specific image. If no objects are detected in an image, the value should be an empty list [].
    Use a 1000x1000 coordinate system for the bounding boxes.

    Example output format for {num_images} images:
    {{
      "image_0": [{{"person": [100, 150, 200, 350]}}, {{"car": [400, 500, 600, 700]}}],
      "image_1": [],
      "image_2": [{{"package": [300, 300, 400, 400]}}]
    }}
    """

        messages = [{"role": "user", "content": image_contents + [{"text": prompt}]}]

        try:
            response = self.bedrock.converse(
                modelId=MODEL_ID,
                messages=messages,
                inferenceConfig={"temperature": 0.0, "maxTokens": 4096, "topP": 1.0},
            )
            model_output = response["output"]["message"]["content"][0]["text"]
            LOGGER.info(f"\n--- {uid}模型对整个批次的原始输出 ---\n{model_output}")

            # 解析模型返回的包含所有图片结果的JSON对象
            batch_results = self.safe_json_load(model_output)
            if not batch_results or not isinstance(batch_results, dict):
                LOGGER.error(f"{uid}模型未返回预期的字典格式结果。")
                return {}

            # --- 核心逻辑：将批处理结果映射回您的格式 ---
            final_output_dict = {}
            for i in range(len(base64_images)):
                # 从批处理结果中获取当前图片的数据，如果不存在则默认为空列表
                nova_detections = batch_results.get(f"image_{i}", [])

                # 转换为您最终需要的格式
                detailed_results = self.format_and_convert_detections(nova_detections)
                final_output_dict[f"file_{i}"] = detailed_results

            return final_output_dict

        except Exception as e:
            LOGGER.error(f"{uid}调用Bedrock模型或处理过程中发生错误: {repr(e)}")
            return {}