Abs-CAM: a gradient optimization interpretable approach for explanation of convolutional neural networks

The black-box nature of deep neural networks severely hinders its performance improvement and application in specific scenes. In recent years, class activation mapping-based method has been widely used to interpret the internal decisions models computer vision tasks. However, when this uses backpropagation obtain gradients, it will cause noise saliency map even locate features that are irrelevant decisions. paper, we propose an absolute value (Abs-CAM) method, which optimizes gradients derived from turns all them into positive enhance visual output neurons’ improve localization ability map. framework Abs-CAM is divided two phases: generating initial final first phase improves by optimizing gradient, second linearly combines with original image semantic information We conduct qualitative quantitative evaluation proposed including Deletion, Insertion, Pointing Game. experimental results show can obviously eliminate map, better related decisions, superior previous methods recognition