Metrics API Reference

Quality Metrics

PSNR (Peak Signal-to-Noise Ratio)

def calculate_psnr(pred: Tensor, target: Tensor, max_val: float = 1.0) -> float

Higher is better. Measures reconstruction quality in dB.

Typical Values: - 20-25 dB: Poor - 25-30 dB: Fair - 30-35 dB: Good - 35+ dB: Excellent

SSIM (Structural Similarity Index)

def calculate_ssim(pred: Tensor, target: Tensor) -> float

Range: [0, 1], where 1 = identical.

Measures structural similarity considering luminance, contrast, and structure.

SAM (Spectral Angle Mapper)

def calculate_sam(pred: Tensor, target: Tensor) -> float

Lower is better (degrees). Measures spectral distortion.

Typical Values: - < 3: Excellent spectral preservation - 3-5: Good - 5-10: Fair - > 10: Poor

ERGAS (Relative Global Error)

def calculate_ergas(pred: Tensor, target: Tensor, ratio: int = 4) -> float

Lower is better. Comprehensive quality metric.

---

Usage

Calculate Individual Metrics

from utils.metrics import calculate_psnr, calculate_ssim, calculate_sam, calculate_ergas

psnr = calculate_psnr(pred, target)
ssim = calculate_ssim(pred, target)
sam = calculate_sam(pred, target)
ergas = calculate_ergas(pred, target)

print(f"PSNR: {psnr:.2f} dB")
print(f"SSIM: {ssim:.4f}")
print(f"SAM: {sam:.4f}")
print(f"ERGAS: {ergas:.4f}")

Calculate All Metrics

from utils.metrics import calculate_metrics

metrics = calculate_metrics(pred, target)
# Returns: {'psnr': ..., 'ssim': ..., 'sam': ..., 'ergas': ...}

---

Metric Interpretation

Metric	Measures	Ideal	Good Range
PSNR	Reconstruction	Higher	> 30 dB
SSIM	Structure	1.0	> 0.85
SAM	Spectral	0	< 5
ERGAS	Overall	0	< 5

Trade-offs

Improving spatial quality (PSNR, SSIM) may degrade spectral quality (SAM), and vice versa. Use appropriate loss functions to balance these trade-offs.