Design and implementation of image based object recognition

Abstract

The aim of this paper is to design and implement image based object recognition. This represents more of a challenge when speaking of advance object recognition systems. A practical example of this issue is the recognition of objects in images. This is a task that humans can perform very well, but convolutional neural network systems don’t struggle to perform. AlexNet pre-trained model was used for the training the dataset because of it trouble-free architecture on very large scale dataset “Cifar-10” using R2019a Matlab. The dataset was split with the ratio of 70% for training and 30% for the testing part.  This has prompted convolutional neural network to start experimenting with networks architectures as well as new algorithms to train them. This research paper presents an approach to train networks such as to improve their robustness to the recognition of object images on R2019a Matlab. This training strategy is then evaluated for designed AlexNet network architecture. The result of the study was that the training algorithm could improve robustness to different image recognition at the expense of an increase in performance for the performance of images of objects (i.e. Dog, Frog, Deer, Automobile, Airplane etc.) with high accuracy of recognition. When the advantages of different types of architectures were evaluated, it was found that accuracy of all object recognition were around 98% based on the image. It followed the findings from classical object recognition that feed-forward neural networks could perform as well their high accuracy of recognition.