A Comprehensive Analytic Scheme for Classification of Novel Models
D. Dakshayani Himabindu, Sudhir Kumar · 2020
Classification is one of the major tasks in supervised learning. This had a great impact on deep leaning in developing various State-of-the-art architectures every year. These models can capture invariances by extracting the finest features by variant design models developed using Convolution Neural Networks. These deep networks have proven their resilience in the large scale image recognition by overhauling the existing drawbacks in architectures.In this research, a complete analytical scheme is developed to extract features from various models and to depict its performance on Fruits-360 dataset. During the analysis, the data is processed from feature extraction block to classification block. After the classification, the individual models' performance is evaluated and assessed with standard classification metrics. Out of these, an optimal model is selected in all prospects. It is observed that VGG models tend to outperform the existing models by capturing invariances by providing transferability on variant tasks. The networks VGG-16, VGG-19 obtained the highest precision of 97.2±0.005, 96.6±0.66 respectively. A pair-plot is obtained to analyse the behaviour pattern of predicted labels to that of true labels. This research is a first attempt to analyse the behaviour of pre-trained models on Fruits-360 dataset.