#码力全开·技术π对#在使用Google Vision API时，如何优化图像识别的准确率？

from google.cloud import vision
import cv2
import numpy as np

def preprocess_image(image_path):
    # 使用OpenCV进行预处理
    img = cv2.imread(image_path)
    
    # 1. 自适应直方图均衡化（CLAHE）
    lab = cv2.cvtColor(img, cv2.COLOR_BGR2LAB)
    l, a, b = cv2.split(lab)
    clahe = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8,8))
    l_clahe = clahe.apply(l)
    lab = cv2.merge((l_clahe, a, b))
    img = cv2.cvtColor(lab, cv2.COLOR_LAB2BGR)
    
    # 2. 降噪（非局部均值去噪）
    img = cv2.fastNlMeansDenoisingColored(img, None, 10, 10, 7, 21)
    
    # 3. 分辨率提升（ESPCN超分辨率）
    sr = cv2.dnn_superres.DnnSuperResImpl_create()
    sr.readModel('ESPCN_x2.pb')
    sr.setModel('espcn', 2)
    img = sr.upsample(img)
    
    return img

# 调用API
client = vision.ImageAnnotatorClient()
processed_img = preprocess_image('input.jpg')
_, encoded_img = cv2.imencode('.jpg', processed_img)
image = vision.Image(content=encoded_img.tobytes())

# 1. 特定特征增强
features = [
    {"type_": vision.Feature.Type.LABEL_DETECTION, "max_results": 50},
    {"type_": vision.Feature.Type.OBJECT_LOCALIZATION, "model": "builtin/latest"}
]

# 2. 上下文提示（提升特定场景识别）
image_context = vision.ImageContext(
    language_hints=["en", "zh"],
    crop_hints_params={
        "aspect_ratios": [0.8, 1.0, 1.2]  # 针对不同长宽比优化
    }
)

response = client.annotate_image({
    "image": image,
    "features": features,
    "image_context": image_context
})

# 1. 多模型投票机制
def ensemble_results(api_responses):
    from collections import Counter
    all_labels = []
    for resp in api_responses:
        all_labels.extend([label.description for label in resp.label_annotations])
    
    # 取Top3高频结果
    return [item[0] for item in Counter(all_labels).most_common(3)]

# 2. 置信度过滤
valid_objects = [
    obj for obj in response.localized_object_annotations 
    if obj.score >= 0.7  # 只保留高置信度结果
]

# 使用AutoML Vision训练补充模型
from google.cloud import automl

# 上传领域特定数据
dataset_client = automl.AutoMlClient()
dataset = {
    "display_name": "medical_images_dataset",
    "image_classification_dataset_metadata": {
        "classification_type": "MULTICLASS"
    }
}
operation = dataset_client.create_dataset(parent="projects/your-project", dataset=dataset)

# 混合Google Vision API和本地模型结果
def hybrid_inference(image):
    api_result = client.annotate_image(...)
    local_model_result = torch_model.predict(image)
    
    # 加权融合（API权重0.7，本地模型0.3）
    final_score = 0.7*api_result.score + 0.3*local_model_result.confidence
    return final_score

def low_light_enhancement(img):
    # 使用Retinex算法
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    blur = cv2.GaussianBlur(gray, (0,0), 3)
    retinex = cv2.addWeighted(gray, 1.5, blur, -0.5, 10)
    return cv2.cvtColor(retinex, cv2.COLOR_GRAY2BGR)