Global Maximal Pattern and Hybrid Optimization-Driven Deep Learning Framework for Object Detection and Recognition

Abstract

Object detection and recognition is the most fundamental, but challenging in the field of computer vision. Object detection identifies the presence of various individual objects in an image. However, these methods suffer from several issues due to scale variations, cluttered background and different orientations. To solve these issues, this paper presents a novel method for increasing the performance of object detection using the proposed atom-exponential weighed moving average (A-EWMA) optimization. The A-EWMA is designed by integrating atom search optimization (ASO) and Exponential Weighed Moving Average (EWMA). Initially, the input image is pre-processed and the pre-processed image is fed to the object detection module, where the object is detected using bounding box segmentation. After that, the hierarchical skeleton and the proposed global maximal pattern features are utilized for extracting the features. The hierarchical skeleton iterates itself to remove all the skeleton branches corresponding to the unimportant shape regions depending on the boundary extraction method, and the proposed global maximal pattern is the new kind of descriptor to address the short-coming of the rotative maximal pattern. The resultant extracted features are employed in object recognition process for recognizing the objects.  In order to facilitate effective training, the Deep Regression Neural network (Deep RNN) is employed wherein, the training of Deep RNN is performed using A-EWMA algorithm. Thus, the objects from the input images are effectively detected using the Deep RNN, which is trained using the proposed A-EWMA algorithm to boost the performance of object detection mechanism. The effectiveness of proposed A-EWMA-based DeepRNN is computed which revealed maximal accuracy of 0.891, Sensitivity of 0.872, and specificity of 0.914.