Whale optimization based approach to compress and fasten CNN for crop disease and species identification

Abstract

In recent years deep learning and machine learning have been widely researched for image based recognition. This research proposes a simplified CNN with 3 layers for classification from 39 classes of crops and their diseases. It also evaluates the performance of pre-trained models such as VGG16 and ResNet50 using transfer learning. Similarly traditional Machine Learning algorithms have been trained and tested on the same dataset. The best accuracy using proposed CNN was 87.67% whereas VGG16 gave best accuracy of 91.51% among Convolution Neural Network models. Similarly Random Forest machine learning method gave best accuracy of 93.02% among Machine Learning models. Since the pre-trained models are having huge size hence in order to deploy these solutions on tiny edge devices compression is done using Whale Optimization. The maximum compesssion was obtained with VGG16 of 88.19% without loss in any performance. It also helped betterment of inference time of 44.13% for proposed CNN, 56.76% for VGG16 and 63.23% for ResNet50.