Performance Analysis Of Several Convolutional Neural Network Models For Image Classification

CNN is a biologically inspired feedforward network with sizeable computational complexity.The advantage of CNN lies in processing large amounts of image data with fewer parameters than an artificial neural network(ANN).Deep neural networks have become a promising solution for numerous machine learning tasks,including image classification,object detection,weather forecasting,etc.Indeed,a lot of papers on CNN-based image classification revealed some error rates.Most of those works have used multiple models and struggled to have a direct path to efficient results.We,through the present research,wish to ease the workout when choosing the best model.The project main objective is(1)the Practical implementation of deep neural network models using VGG16,Resnet 50,Inception V3,and Xception.(2)Compare and evaluate all the above model 's performance in accuracy(3)Layout an in-depth insight into a comprehensive work with CNN's models.In this project,we used a plant seedling dataset with 12 categories with different numbers of images.Our experiment has divided into three parts.(1)Comparing all the above model's performance without fine-tuning on the same learning rate 0.0001 and the same optimizer(Adam)for all above models on the plant seedling dataset using Image Net pre-trained models directly downloaded from the Keras application.(2)Comparing the above accuracy of the models with fine-tuning on the same optimizer(Adam)and learning rate 0.0001 with Image Net pre-trained models,downloaded directly from the Keras,all the models operated on the plant seedling dataset.(3)We compare batch size effect in accuracy in models training on the same learning rate,which is 0.0001 using the Adam optimizer on the plant seedling data set.Furthermore,in the comparison,we found the Res Net50 without fine-tuning with 88 % best testing accuracy among the proposed models and the Xception model with fine-tuning 97% best accuracy for the plant seedling dataset.We have also experimentally proved that the training and testing accuracy can increase by increasing the batch size in model training.