Enhancing Deep Neural Network Saliency Visualizations with Gradual Extrapolation

We propose an enhancement technique of the Class Activation Mapping methods like Grad-CAM or Excitation Backpropagation, which presents visual explanations of decisions from CNN-based models. Our idea, called Gradual Extrapolation, can supplement any method that generates a heatmap picture by sharpening the output. Instead of producing a coarse localization map highlighting the important predictive regions in the image, our method outputs the specific shape that most contributes to the model output. Thus, it improves the accuracy of saliency maps. Effect has been achieved by gradual propagation of the crude map obtained in deep layer through all preceding layers with respect to their activations. In validation tests conducted on a selected set of images, the proposed method significantly improved the localization detection of the neural networks' attention. Furthermore, the proposed method is applicable to any deep neural network model.