High Performance Classification of Caltech-101 with a Transfer Learned Deep Convolutional Neural Network

Numerous models and working schemes have been proposed through decades for the successful recognition of the objects. Significant contributions are notable in the field of object recognition. However, near accurate recognition is still a challenge in this domain. In this research, we considered the Caltech-101 dataset having 102 diverse and imbalanced classes i.e., people, animals, landscapes, structures, furniture, etc. which made the recognition more complicated. We proposed and utilized modified InceptionV3 and modified EfficientNetB6 architectures for the recognition of objects which obtained 99.65% and 99.72% overall accuracy respectively. We further showed via experimental analysis that the softmax-averaging technique can further boost the accuracy to 99.85% and all three proposed procedures suppressed the previous studies by a notable boundary as well.