Métricas de Calidad de Imagen IA LPIPS・SSIM Guía Práctica 2025

import torch
import torch.nn as nn
import lpips
from torchvision import models, transforms

class LPIPSEvaluator:
    def __init__(self, net='alex', use_gpu=True):
        """
        Inicialización del modelo LPIPS
        net: Elegir de 'alex', 'vgg', 'squeeze'
        """
        self.loss_fn = lpips.LPIPS(net=net)
        self.device = torch.device('cuda' if use_gpu and torch.cuda.is_available() else 'cpu')
        self.loss_fn.to(self.device)
        
        # Pipeline de preprocesamiento
        self.transform = transforms.Compose([
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406],
                               std=[0.229, 0.224, 0.225])
        ])
    
    def calculate_lpips(self, img1, img2):
        """
        Calcular distancia LPIPS entre dos imágenes
        """
        # Preprocesamiento
        tensor1 = self.transform(img1).unsqueeze(0).to(self.device)
        tensor2 = self.transform(img2).unsqueeze(0).to(self.device)
        
        # Cálculo LPIPS
        with torch.no_grad():
            distance = self.loss_fn(tensor1, tensor2)
        
        return distance.item()
    
    def batch_evaluate(self, image_pairs):
        """
        Evaluación LPIPS con procesamiento por lotes
        """
        results = []
        
        for img1, img2 in image_pairs:
            lpips_score = self.calculate_lpips(img1, img2)
            results.append({
                'lpips_distance': lpips_score,
                'perceptual_similarity': 1 - lpips_score,  # Expresar como similitud
                'quality_category': self.categorize_quality(lpips_score)
            })
        
        return results
    
    def categorize_quality(self, lpips_score):
        """
        Clasificación de categoría de calidad basada en puntuación LPIPS
        """
        if lpips_score < 0.1:
            return 'excellent'
        elif lpips_score < 0.2:
            return 'good'
        elif lpips_score < 0.4:
            return 'acceptable'
        else:
            return 'poor'

class CustomLPIPSNetwork(nn.Module):
    def __init__(self, backbone='resnet50'):
        super().__init__()
        
        # Selección de red backbone
        if backbone == 'resnet50':
            self.features = models.resnet50(pretrained=True)
            self.features = nn.Sequential(*list(self.features.children())[:-2])
        elif backbone == 'efficientnet':
            self.features = models.efficientnet_b0(pretrained=True).features
        
        # Capas de extracción de características
        self.feature_layers = [1, 4, 8, 12, 16]  # Índices de capas a extraer
        
        # Capas de transformación lineal
        self.linear_layers = nn.ModuleList([
            nn.Sequential(
                nn.Conv2d(64, 1, 1, bias=False),
                nn.GroupNorm(1, 1, affine=False)
            ),
            nn.Sequential(
                nn.Conv2d(256, 1, 1, bias=False),
                nn.GroupNorm(1, 1, affine=False)
            ),
            nn.Sequential(
                nn.Conv2d(512, 1, 1, bias=False),
                nn.GroupNorm(1, 1, affine=False)
            )
        ])
    
    def forward(self, x1, x2):
        # Extracción de características
        features1 = self.extract_features(x1)
        features2 = self.extract_features(x2)
        
        # Cálculo de distancia en cada capa
        distances = []
        for i, (f1, f2) in enumerate(zip(features1, features2)):
            # Normalización L2
            f1_norm = f1 / (torch.norm(f1, dim=1, keepdim=True) + 1e-8)
            f2_norm = f2 / (torch.norm(f2, dim=1, keepdim=True) + 1e-8)
            
            # Cálculo de distancia
            diff = (f1_norm - f2_norm) ** 2
            
            # Transformación lineal
            if i < len(self.linear_layers):
                diff = self.linear_layers[i](diff)
            
            # Promedio espacial
            distance = torch.mean(diff, dim=[2, 3])
            distances.append(distance)
        
        # Promedio ponderado
        total_distance = sum(distances) / len(distances)
        return total_distance

import numpy as np
from skimage.metrics import structural_similarity
from scipy.ndimage import gaussian_filter

class SSIMEvaluator:
    def __init__(self, window_size=11, k1=0.01, k2=0.03, sigma=1.5):
        self.window_size = window_size
        self.k1 = k1
        self.k2 = k2
        self.sigma = sigma
    
    def calculate_ssim(self, img1, img2, data_range=1.0):
        """
        Cálculo SSIM básico
        """
        return structural_similarity(
            img1, img2,
            data_range=data_range,
            multichannel=True,
            gaussian_weights=True,
            sigma=self.sigma,
            use_sample_covariance=False
        )
    
    def calculate_ms_ssim(self, img1, img2, weights=None):
        """
        Implementación Multi-Scale SSIM (MS-SSIM)
        """
        if weights is None:
            weights = [0.0448, 0.2856, 0.3001, 0.2363, 0.1333]
        
        levels = len(weights)
        mssim = 1.0
        
        for i in range(levels):
            ssim_val = self.calculate_ssim(img1, img2)
            
            if i < levels - 1:
                # Submuestreo
                img1 = self.downsample(img1)
                img2 = self.downsample(img2)
                mssim *= ssim_val ** weights[i]
            else:
                mssim *= ssim_val ** weights[i]
        
        return mssim
    
    def downsample(self, img):
        """
        Filtrado Gaussiano + submuestreo
        """
        filtered = gaussian_filter(img, sigma=1.0, axes=[0, 1])
        return filtered[::2, ::2]
    
    def ssim_map(self, img1, img2):
        """
        Generar mapa SSIM
        """
        # Conversión a escala de grises
        if len(img1.shape) == 3:
            img1_gray = np.mean(img1, axis=2)
            img2_gray = np.mean(img2, axis=2)
        else:
            img1_gray = img1
            img2_gray = img2
        
        # Media
        mu1 = gaussian_filter(img1_gray, self.sigma)
        mu2 = gaussian_filter(img2_gray, self.sigma)
        
        mu1_sq = mu1 ** 2
        mu2_sq = mu2 ** 2
        mu1_mu2 = mu1 * mu2
        
        # Varianza y covarianza
        sigma1_sq = gaussian_filter(img1_gray ** 2, self.sigma) - mu1_sq
        sigma2_sq = gaussian_filter(img2_gray ** 2, self.sigma) - mu2_sq
        sigma12 = gaussian_filter(img1_gray * img2_gray, self.sigma) - mu1_mu2
        
        # Cálculo SSIM
        c1 = (self.k1 * 1.0) ** 2
        c2 = (self.k2 * 1.0) ** 2
        
        ssim_map = ((2 * mu1_mu2 + c1) * (2 * sigma12 + c2)) / \
                   ((mu1_sq + mu2_sq + c1) * (sigma1_sq + sigma2_sq + c2))
        
        return ssim_map

import torch
import torchvision.models as models

class DISTSEvaluator:
    def __init__(self, use_gpu=True):
        self.device = torch.device('cuda' if use_gpu and torch.cuda.is_available() else 'cpu')
        
        # Usar parte de extracción de características de red VGG
        vgg = models.vgg16(pretrained=True).features
        self.stages = nn.ModuleList([
            vgg[:4],   # conv1_2
            vgg[:9],   # conv2_2
            vgg[:16],  # conv3_3
            vgg[:23],  # conv4_3
            vgg[:30]   # conv5_3
        ]).to(self.device)
        
        for param in self.stages.parameters():
            param.requires_grad = False
    
    def extract_features(self, x):
        features = []
        for stage in self.stages:
            x = stage(x)
            features.append(x)
        return features
    
    def calculate_dists(self, img1, img2):
        """
        Calcular DISTS (Deep Image Structure and Texture Similarity)
        """
        # Preprocesamiento
        tensor1 = self.preprocess(img1).to(self.device)
        tensor2 = self.preprocess(img2).to(self.device)
        
        # Extracción de características
        feats1 = self.extract_features(tensor1)
        feats2 = self.extract_features(tensor2)
        
        structure_score = 0
        texture_score = 0
        
        for f1, f2 in zip(feats1, feats2):
            # Similitud de estructura (similitud de media)
            struct_sim = self.structure_similarity(f1, f2)
            structure_score += struct_sim
            
            # Similitud de textura (similitud de covarianza)
            texture_sim = self.texture_similarity(f1, f2)
            texture_score += texture_sim
        
        # Composición ponderada
        alpha = 0.8  # peso estructura
        beta = 0.2   # peso textura
        
        dists_score = alpha * structure_score + beta * texture_score
        return dists_score.item()
    
    def structure_similarity(self, feat1, feat2):
        """
        Calcular similitud de estructura
        """
        # Media por dirección de canal
        mean1 = torch.mean(feat1, dim=1, keepdim=True)
        mean2 = torch.mean(feat2, dim=1, keepdim=True)
        
        # Similitud estructural
        numerator = 2 * mean1 * mean2
        denominator = mean1 ** 2 + mean2 ** 2
        
        structure_map = numerator / (denominator + 1e-8)
        return torch.mean(structure_map)
    
    def texture_similarity(self, feat1, feat2):
        """
        Calcular similitud de textura
        """
        # Calcular matriz de covarianza de mapas de características
        b, c, h, w = feat1.shape
        feat1_flat = feat1.view(b, c, -1)
        feat2_flat = feat2.view(b, c, -1)
        
        # Cálculo de covarianza
        cov1 = torch.bmm(feat1_flat, feat1_flat.transpose(1, 2)) / (h * w - 1)
        cov2 = torch.bmm(feat2_flat, feat2_flat.transpose(1, 2)) / (h * w - 1)
        
        # Similitud por norma de Frobenius
        diff_norm = torch.norm(cov1 - cov2, 'fro', dim=[1, 2])
        max_norm = torch.maximum(torch.norm(cov1, 'fro', dim=[1, 2]),
                                torch.norm(cov2, 'fro', dim=[1, 2]))
        
        texture_sim = 1 - diff_norm / (max_norm + 1e-8)
        return torch.mean(texture_sim)

from scipy.linalg import sqrtm
import numpy as np

class FIDEvaluator:
    def __init__(self):
        # Modelo Inception v3 (para extracción de características)
        self.inception = models.inception_v3(pretrained=True, transform_input=False)
        self.inception.fc = nn.Identity()  # Remover capa de clasificación
        self.inception.eval()
        
        for param in self.inception.parameters():
            param.requires_grad = False
    
    def extract_features(self, images):
        """
        Extracción de características usando Inception v3
        """
        features = []
        
        with torch.no_grad():
            for img in images:
                # Redimensionar a tamaño apropiado (299x299)
                img_resized = F.interpolate(img.unsqueeze(0), 
                                          size=(299, 299), 
                                          mode='bilinear')
                
                feat = self.inception(img_resized)
                features.append(feat.cpu().numpy())
        
        return np.concatenate(features, axis=0)
    
    def calculate_fid(self, real_images, generated_images):
        """
        Calcular FID (Fréchet Inception Distance)
        """
        # Extracción de características
        real_features = self.extract_features(real_images)
        gen_features = self.extract_features(generated_images)
        
        # Cálculo de estadísticas
        mu_real = np.mean(real_features, axis=0)
        sigma_real = np.cov(real_features, rowvar=False)
        
        mu_gen = np.mean(gen_features, axis=0)
        sigma_gen = np.cov(gen_features, rowvar=False)
        
        # Cálculo de distancia de Fréchet
        diff = mu_real - mu_gen
        covmean = sqrtm(sigma_real.dot(sigma_gen))
        
        # Remover componentes imaginarios debido a errores numéricos
        if np.iscomplexobj(covmean):
            covmean = covmean.real
        
        fid = diff.dot(diff) + np.trace(sigma_real + sigma_gen - 2 * covmean)
        
        return fid

class ComprehensiveQualityEvaluator:
    def __init__(self):
        self.lpips_evaluator = LPIPSEvaluator()
        self.ssim_evaluator = SSIMEvaluator()
        self.dists_evaluator = DISTSEvaluator()
        self.fid_evaluator = FIDEvaluator()
        
        # Configuración de pesos
        self.weights = {
            'lpips': 0.3,
            'ssim': 0.3,
            'dists': 0.2,
            'psnr': 0.1,
            'fid': 0.1
        }
    
    def evaluate_single_pair(self, img1, img2):
        """
        Evaluación de calidad integral de par de imágenes
        """
        results = {}
        
        # LPIPS
        results['lpips'] = self.lpips_evaluator.calculate_lpips(img1, img2)
        
        # SSIM
        results['ssim'] = self.ssim_evaluator.calculate_ssim(img1, img2)
        
        # DISTS
        results['dists'] = self.dists_evaluator.calculate_dists(img1, img2)
        
        # PSNR (valor de referencia)
        results['psnr'] = self.calculate_psnr(img1, img2)
        
        # Calcular puntuación compuesta
        composite_score = self.calculate_composite_score(results)
        results['composite_score'] = composite_score
        
        # Determinar nivel de calidad
        results['quality_level'] = self.determine_quality_level(composite_score)
        
        return results
    
    def calculate_psnr(self, img1, img2):
        """
        Cálculo PSNR
        """
        mse = np.mean((img1 - img2) ** 2)
        if mse == 0:
            return float('inf')
        return 20 * np.log10(1.0 / np.sqrt(mse))
    
    def calculate_composite_score(self, metrics):
        """
        Puntuación compuesta de múltiples métricas
        """
        # Normalizar cada métrica al rango 0-1
        normalized_scores = {
            'lpips': 1 - min(metrics['lpips'], 1.0),  # Menor es mejor
            'ssim': metrics['ssim'],                   # Mayor es mejor
            'dists': metrics['dists'],                 # Mayor es mejor
            'psnr': min(metrics['psnr'] / 50, 1.0),   # Normalización
        }
        
        # Composición ponderada
        composite = sum(
            self.weights[metric] * score 
            for metric, score in normalized_scores.items()
            if metric in self.weights
        )
        
        return composite
    
    def determine_quality_level(self, score):
        """
        Determinación de nivel de calidad basado en puntuación
        """
        if score >= 0.9:
            return 'excellent'
        elif score >= 0.8:
            return 'very_good'
        elif score >= 0.7:
            return 'good'
        elif score >= 0.6:
            return 'acceptable'
        elif score >= 0.5:
            return 'poor'
        else:
            return 'very_poor'

import asyncio
import aiofiles
from pathlib import Path

class BatchQualityEvaluator:
    def __init__(self, evaluator, max_workers=4):
        self.evaluator = evaluator
        self.max_workers = max_workers
        self.semaphore = asyncio.Semaphore(max_workers)
    
    async def evaluate_directory(self, original_dir, processed_dir, output_file):
        """
        Evaluación por lotes por directorio
        """
        original_path = Path(original_dir)
        processed_path = Path(processed_dir)
        
        # Obtener pares de archivos de imagen
        image_pairs = self.get_image_pairs(original_path, processed_path)
        
        # Evaluación por lotes con procesamiento paralelo
        tasks = [
            self.evaluate_pair_async(orig, proc) 
            for orig, proc in image_pairs
        ]
        
        results = await asyncio.gather(*tasks, return_exceptions=True)
        
        # Generar reporte
        report = self.generate_report(image_pairs, results)
        
        # Guardar resultados
        await self.save_report(report, output_file)
        
        return report
    
    async def evaluate_pair_async(self, original_path, processed_path):
        """
        Evaluación asíncrona de par de imágenes
        """
        async with self.semaphore:
            # Cargar imágenes
            img1 = await self.load_image_async(original_path)
            img2 = await self.load_image_async(processed_path)
            
            # Ejecutar evaluación
            result = self.evaluator.evaluate_single_pair(img1, img2)
            result['original_path'] = str(original_path)
            result['processed_path'] = str(processed_path)
            
            return result
    
    async def load_image_async(self, path):
        """
        Carga asíncrona de imagen
        """
        async with aiofiles.open(path, 'rb') as f:
            data = await f.read()
        
        # Decodificar imagen con PIL
        from PIL import Image
        import io
        img = Image.open(io.BytesIO(data))
        return np.array(img) / 255.0
    
    def generate_report(self, image_pairs, results):
        """
        Generar reporte de evaluación
        """
        successful_results = [r for r in results if not isinstance(r, Exception)]
        
        # Cálculo de estadísticas
        stats = {
            'total_images': len(image_pairs),
            'successful_evaluations': len(successful_results),
            'average_composite_score': np.mean([r['composite_score'] for r in successful_results]),
            'average_lpips': np.mean([r['lpips'] for r in successful_results]),
            'average_ssim': np.mean([r['ssim'] for r in successful_results]),
            'quality_distribution': self.calculate_quality_distribution(successful_results)
        }
        
        report = {
            'summary': stats,
            'detailed_results': successful_results,
            'failed_evaluations': [r for r in results if isinstance(r, Exception)]
        }
        
        return report
    
    async def save_report(self, report, output_file):
        """
        Guardar reporte como JSON
        """
        import json
        async with aiofiles.open(output_file, 'w') as f:
            await f.write(json.dumps(report, indent=2, default=str))

import threading
import queue
from collections import deque

class RealTimeQualityMonitor:
    def __init__(self, evaluator, window_size=100):
        self.evaluator = evaluator
        self.window_size = window_size
        self.quality_history = deque(maxlen=window_size)
        self.alert_queue = queue.Queue()
        self.is_running = False
        
        # Umbrales de alerta
        self.thresholds = {
            'composite_score': {
                'warning': 0.6,
                'critical': 0.4
            },
            'lpips': {
                'warning': 0.3,
                'critical': 0.5
            }
        }
    
    def start_monitoring(self, input_queue):
        """
        Iniciar monitoreo en tiempo real
        """
        self.is_running = True
        monitor_thread = threading.Thread(
            target=self.monitor_loop, 
            args=(input_queue,)
        )
        monitor_thread.start()
        return monitor_thread
    
    def monitor_loop(self, input_queue):
        """
        Bucle principal de monitoreo
        """
        while self.is_running:
            try:
                # Obtener par de imágenes de la cola
                img_pair = input_queue.get(timeout=1.0)
                
                if img_pair is None:  # Señal de terminación
                    break
                
                # Evaluación de calidad
                result = self.evaluator.evaluate_single_pair(*img_pair)
                
                # Agregar al historial
                self.quality_history.append(result)
                
                # Verificar alertas
                self.check_alerts(result)
                
                # Actualizar estadísticas
                self.update_statistics()
                
            except queue.Empty:
                continue
            except Exception as e:
                print(f"Error de monitoreo: {e}")
    
    def check_alerts(self, result):
        """
        Verificar condiciones de alerta
        """
        for metric, thresholds in self.thresholds.items():
            if metric in result:
                value = result[metric]
                
                if value < thresholds['critical']:
                    self.alert_queue.put({
                        'level': 'critical',
                        'metric': metric,
                        'value': value,
                        'threshold': thresholds['critical'],
                        'timestamp': time.time()
                    })
                elif value < thresholds['warning']:
                    self.alert_queue.put({
                        'level': 'warning',
                        'metric': metric,
                        'value': value,
                        'threshold': thresholds['warning'],
                        'timestamp': time.time()
                    })
    
    def get_current_statistics(self):
        """
        Obtener estadísticas actuales
        """
        if not self.quality_history:
            return {}
        
        recent_scores = [r['composite_score'] for r in self.quality_history]
        recent_lpips = [r['lpips'] for r in self.quality_history]
        
        return {
            'window_size': len(self.quality_history),
            'average_quality': np.mean(recent_scores),
            'quality_trend': self.calculate_trend(recent_scores),
            'average_lpips': np.mean(recent_lpips),
            'quality_stability': np.std(recent_scores)
        }

class AdaptiveQualityOptimizer:
    def __init__(self, evaluator, target_quality=0.8):
        self.evaluator = evaluator
        self.target_quality = target_quality
        self.parameter_history = []
        
        # Parámetros objetivo para optimización
        self.parameters = {
            'compression_quality': {'min': 50, 'max': 100, 'current': 85},
            'resize_algorithm': {'options': ['lanczos', 'bicubic', 'bilinear'], 'current': 'lanczos'},
            'sharpening_strength': {'min': 0.0, 'max': 2.0, 'current': 1.0}
        }
    
    def optimize_parameters(self, test_images, max_iterations=50):
        """
        Optimización de parámetros hacia objetivo de calidad
        """
        best_params = self.parameters.copy()
        best_quality = 0
        
        for iteration in range(max_iterations):
            # Procesar con parámetros actuales
            processed_images = self.process_with_parameters(
                test_images, self.parameters
            )
            
            # Evaluación de calidad
            avg_quality = self.evaluate_batch_quality(
                test_images, processed_images
            )
            
            print(f"Iteración {iteration + 1}: Calidad = {avg_quality:.3f}")
            
            # Actualizar mejor resultado
            if avg_quality > best_quality:
                best_quality = avg_quality
                best_params = self.parameters.copy()
            
            # Verificar logro de objetivo
            if avg_quality >= self.target_quality:
                print(f"¡Calidad objetivo {self.target_quality} lograda!")
                break
            
            # Actualizar parámetros
            self.update_parameters(avg_quality)
            
            # Registrar historial
            self.parameter_history.append({
                'iteration': iteration,
                'parameters': self.parameters.copy(),
                'quality': avg_quality
            })
        
        return best_params, best_quality
    
    def update_parameters(self, current_quality):
        """
        Actualización de parámetros basada en calidad actual
        """
        quality_gap = self.target_quality - current_quality
        
        # Usar configuraciones más conservadoras cuando la calidad es baja
        if quality_gap > 0.1:
            # Incrementar calidad de compresión
            self.parameters['compression_quality']['current'] = min(
                100, 
                self.parameters['compression_quality']['current'] + 5
            )
            
            # Reducir nitidez
            self.parameters['sharpening_strength']['current'] = max(
                0.0,
                self.parameters['sharpening_strength']['current'] - 0.1
            )
        
        # Enfocarse en eficiencia cuando la calidad es suficientemente alta
        elif quality_gap < -0.05:
            self.parameters['compression_quality']['current'] = max(
                50,
                self.parameters['compression_quality']['current'] - 2
            )

FROM pytorch/pytorch:1.9.0-cuda10.2-cudnn7-runtime

WORKDIR /app

# Instalar dependencias
COPY requirements.txt .
RUN pip install -r requirements.txt

# Código de aplicación
COPY src/ ./src/
COPY models/ ./models/

# Punto de entrada
COPY entrypoint.sh .
RUN chmod +x entrypoint.sh

EXPOSE 8080

ENTRYPOINT ["./entrypoint.sh"]

from fastapi import FastAPI, File, UploadFile, HTTPException
from fastapi.responses import JSONResponse
import uvicorn

app = FastAPI(title="API de Evaluación de Calidad de Imagen")

# Evaluador global
quality_evaluator = ComprehensiveQualityEvaluator()

@app.post("/evaluate/single")
async def evaluate_single_image(
    original: UploadFile = File(...),
    processed: UploadFile = File(...)
):
    """
    Evaluación de par de imágenes único
    """
    try:
        # Cargar imágenes
        original_img = await load_upload_image(original)
        processed_img = await load_upload_image(processed)
        
        # Ejecutar evaluación
        result = quality_evaluator.evaluate_single_pair(
            original_img, processed_img
        )
        
        return JSONResponse(content=result)
    
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

@app.post("/evaluate/batch")
async def evaluate_batch_images(
    files: List[UploadFile] = File(...)
):
    """
    Evaluación por lotes
    """
    if len(files) % 2 != 0:
        raise HTTPException(
            status_code=400, 
            detail="Se requiere número par de archivos (pares original + procesado)"
        )
    
    results = []
    for i in range(0, len(files), 2):
        original_img = await load_upload_image(files[i])
        processed_img = await load_upload_image(files[i + 1])
        
        result = quality_evaluator.evaluate_single_pair(
            original_img, processed_img
        )
        results.append(result)
    
    # Cálculo de estadísticas
    summary = {
        'total_pairs': len(results),
        'average_quality': np.mean([r['composite_score'] for r in results]),
        'quality_distribution': calculate_quality_distribution(results)
    }
    
    return JSONResponse(content={
        'summary': summary,
        'results': results
    })

@app.get("/health")
async def health_check():
    return {"status": "healthy"}

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8080)