A Study of Data Processing for Object Recognition in Scene Image using FRCNN: A Smart Grid Technology

Abstract

This paper proposes a new learning about the significance of part elements of a scene image with the effort of data processing. A top down tree structure with every node representing an annotation or bounding box having labeled visual features of an object existing in the image is studied in the paper. The images and its object annotations are from a trained dataset and are parsed to obtain the proposed representation. The images from the datasets and their parsed tree representations will be trained using a network called LSTM (Long Short Term Memory) network. The object detection may not be agnostic to the entire content of the image due to being influenced by the image composition and the discovered parts. The attempt has been made to show the object detection as a representation of the objects and their locations, parts of these objects, and the accuracy of the object detection method has been noted to have an efficient record with the implementation of the baseline Fast Region Based Convolutional Neural Network(FRCNN) method. The tested google open images datasets are used and found to have increased object detection record in notably respect to the use of high cost sensors in digital devices.