[ICCV 2021] Focal Frequency Loss for Image Reconstruction and Synthesis

No description available

Tensorflow Implementation of Focal Frequency Loss for Image Reconstruction and Synthesis [ICCV 2021]

Log Focal Frequency Loss for Bioimage Restoration (accepted IEEE ISBI 2026)

In this study we will introduce atrial fibrillation, one of the cardiac arrhythmias, and see how it can be diagnosed using convolutional neural networks in combination with various methods and supervised learning models, including: gray-level co-occurrence matrix, short-time Fourier transform based spectrogram, support-vector machines, k-nearest neighbors, multi-layer perceptron, focal loss, multi-scale decomposition, time-frequency analysis.

These are the results for Focal Frequency Loss for Image Reconstruction and Synthesis paper but with different dataset. In this case the TFW: Annotated Thermal Faces in the Wild Dataset. By using novel focal frequency loss it allows a model to adaptively focus on frequency components that are hard to synthesize by down weighting the easy ones.

A comprehensive extension of the Focal Loss by taking into account class similarities, frequencies and severity.

Robust 3-class Alzheimer's classification benchmark on OASIS-1 with subject-level splits. Demonstrates the non-viability of the standard 4-class formulation and evaluates WCE, Focal Loss, LDAM, and proposed Frequency-Adaptive Focal Loss (FA-FL) across 5 seeds. EfficientNet-B0 backbone.

Dual-AttnGAN: A dual-input GAN for joint low-light image enhancement and 4× super-resolution, fusing wide-angle and zoomed images via asymmetric cross-attention.