Sound Event Detection on Grad­CAM method with Complex Pretrained Model

Grad­CAM weakly supervised algorithm is usually used in classification tasks as a feature visualization method. This method can mark the pixel points determined by CNN as classification features by the target layer parameters of convolutional neural network. Based on the ability of Grad­CAM to locate classification features, this paper combines the audio event detection task with Grad­CAM algorithm to form an audio event detection system based on Grad­CAM.The feature localization ability obtained by convolutional neural network in the audio event classification task is used to assist in completing the audio event detection task. In the experiment based on DCASE2022task4 dataset, the PSDS2 score of 0.74157 is higher than the baseline system with PSDS2 score of 0.54336.