Mask-GradCAM: Object Identification and Localization of Visual Presentation for Deep Convolutional Network

摘要

This paper presents the conceptually simple, flexible and more suitable framework to demonstrate object localization and object recognition by Mask RCNN along with Grad-CAM (Mask-GradCAM) method that is mainly used to build framework to provide the better visual identification. Because Mask RCNN based method provides a function that take array of pixel values for load the images and aspects of the prediction dictionary such that all class labels, scores, bounding boxes and will plot the image with all these annotations. However, applying this method along with Mask RCNN raises significant challenges such that image pixel issues and quality breaking while producing the heat map. Therefore, this entire system is mentioned as Mask-Grad CAM. This research work considers GradCAM++ and GradCAM methods as basic functionality derivation to implement the Mask-GradCAM. In this combination of MaskRCNN along with Grad-CAM provide the fine-grained display to generate higher resolution visual representation. Hence this method is presented as Mask-GradCAM.