Analysis of Loss Functions for Image Reconstruction Using Convolutional Autoencoder

Nishant Khare, Poornima Singh Thakur, Pritee Khanna, Aparajita Ojha

Abstract: In recent years, several loss functions have been proposed for the image reconstruction task of convolutional autoencoders (CAEs). In this paper, a performance analysis of a CAE with respect to different loss functions is presented. Quality of reconstruction is analyzed using the mean Square error (MSE), binary cross-entropy (BCE), Sobel, Laplacian, and Focal binary loss functions. To evaluate the performance of different loss functions, a vanilla autoencoder is trained on eight datasets having diversity in terms of application domains, image dimension, color space, and the number of images in the dataset. MSE, peak signal to noise ratio (PSNR), and structural similarity index (SSIM) metrics have been used as the performance measures on all eight datasets. The assessment shows that the MSE loss function outperforms two datasets with a small image dimension and a large number of images. At the same time, BCE excels on six datasets with high image dimensions and a small number of training samples in datasets compared with the Sobel and Laplacian loss functions.