Image Fusion Evaluation Metrics

This document presents a curated collection of image fusion evaluation standards comprising 13 performance metrics. These metrics are essential for assessing the effectiveness of image fusion algorithms. Below are detailed explanations with implementation considerations for each metric: 1. Average Gradient: Measures the degree of color variation across different regions in an image. Implementation typically involves calculating gradient magnitude using Sobel or Prewitt operators followed by averaging across all pixels. 2. Edge Intensity: Evaluates the clarity of object edges in fused images. Can be computed using edge detection algorithms like Canny or Sobel, then quantifying edge strength through gradient magnitude summation. 3. Information Entropy: Represents the information distribution within an image. Calculated using the probability distribution of gray levels: H = -∑(p_i * log2(p_i)), where p_i is the probability of gray level i. 4. Gray Mean Value: Computes the average pixel intensity value across the entire image. Simple implementation using mean() function on the image matrix. 5. Standard Deviation (Mean Square Error MSE): Measures the dispersion of pixel intensity values. MSE calculation: MSE = (1/n)∑(I_original - I_fused)², where n is total pixels. 6. Root Mean Square Error: Quantifies the error level in synthesized images. RMSE = sqrt(MSE), commonly used with reference images for quality assessment. 7. Peak Signal-to-Noise Ratio (PSNR): Assesses the signal-to-noise ratio in fused images. PSNR = 20*log10(MAX_I / sqrt(MSE)), where MAX_I is maximum possible pixel value. 8. Spatial Frequency (SF): Indicates the frequency distribution within an image. Computed as SF = sqrt(RF² + CF²), where RF and CF represent row and column frequencies. 9. Image Definition: Evaluates the clarity level of images. Often implemented using Brenner gradient or Tenengrad measure based on Sobel operators. 10. Mutual Information (MI): Measures similarity between two images. MI(X,Y) = H(X) + H(Y) - H(X,Y), where H represents entropy, requiring joint histogram calculation. 11. Structural Similarity (SSIM): Assesses structural similarity in fused images. Implements luminance, contrast, and structure comparison: SSIM = [l(x,y)]^α * [c(x,y)]^β * [s(x,y)]^γ. 12. Cross Entropy: Measures information difference between images. For distributions P and Q: CE = -∑P(i)logQ(i), often used to compare original and fused image histograms. 13. Relative Standard Deviation: Evaluates the variation degree of pixel intensity values. Calculated as (Standard Deviation / Mean) * 100%, providing normalized dispersion measure. We hope this information proves valuable for your image fusion projects. Your contributions and discussions about implementation techniques are highly welcomed!