High-Performance Object Recognition by Employing a Transfer Learned Deep Convolutional Neural Network

Researchers have been conveying and producing various models for the precise identification of objects for a long period now. Distinguished contributions are notable in the field of object identification. Numerous benchmark datasets are now available for composing an effective model that can discover a large number of target classes without a tremendous loss in performance. In this study, we began with a benchmark dataset Caltech-101 which consists of 102 classes. Being an extremely imbalanced dataset and having a blend of RGB and Gray images, the recognition is extremely challenging. We proposed and implemented a modified DenseNet-201 design after the preprocessing steps and produced an overall accuracy of 99.73% which exceeded all the previous results or achievements by a noteworthy boundary.