Object Detection Algorithm Based on Multi-Scaled Convolutional Neural Networks
T J Nandhini, K. Thinakaran
Abstract
Object detection algorithms must first identify all the objects inside an image before machine vision can properly categorize and localize them. Many methods have been proposed to handle this problem, with most of the motivation coming from computer vision and deep learning methods. However, prevailing technologies have never effectively recognized tiny, dense things and often failed to detect objects that have undergone random geometric alterations. We analyze the current state of the art in object identification and propose a deformable convolutional network with adjustable depths to address these concerns. The results of our research suggest that they are better than the current best practices, blend deep convolutional networks with flexible convolutional structures to account for geometric variations, and get multi-scaled features. Next, we perform the remaining phases of object identification and region regress by up-sampling the fusion of multi-scaled elements. Experimental validation of our proposed framework demonstrates a considerable improvement in accuracy relative to time spent recognizing small target objects with geometric distortion.