Robust Object Detection and Localization Using Semantic Segmentation Network

Abstract

The advancements in the area of object localization are in great progress for analyzing the spatial relations of different objects from the set of images. Several object localization techniques rely on classification, which decides, if the object exist or not, but does not provide the object information using pixel-wise segmentation. This work introduces an object detection and localization technique using semantic segmentation network (SSN) and deep convolutional neural network (Deep CNN). Here, the proposed technique consists of the following steps: Initially, the image is denoised using the filtering to eliminate the noise present in the image. Then, pre-processed image undergoes sparking process for making the image suitable for the segmentation using SSN for object segmentation. The obtained segments are subjected as the input to the proposed Stochastic-Cat Crow optimization (Stochastic-CCO)-based Deep CNN for the object classification. Here, the proposed Stochastic-CCO, obtained by integrating stochastic gradient descent and the CCO, is used for training the Deep CNN. The CCO is designed by the integration of cat swarm optimization (CSO) and crow search algorithm and takes advantages of both optimization algorithms. The experimentation proves that the proposed Stochastic-CCO-based Deep CNN-based technique acquired maximal accuracy of 98.7.