vioft2nntf2t|tblJournal|Abstract_paper|0xf4ff17362c000000af7e0b0001000200 Convolutional Neural Network (CNN), a particular type of forwarding feed network composed of convolutional, pooling, and fully connected layers, has become the dominant and most widely used deep learning architecture. Significantly enhanced effectiveness of ConvNets has made CNNs the go-to architecture model for almost every image processing-based application. CNNs automatically and adaptively learn spatial hierarchies of features with high accuracy, precision, and efficiency. This paper proposes three CNN models with 5, 6, and 7 layers with two types of classification layers at the top of the model, resulting in six kinds of models. Each model is trained on apple leaf diseases obtained with augmentation deployed on the PlantVillage dataset containing images of healthy and three types of leaf diseases. The trained models are compared on training time, testing accuracy, testing time. The best performing model (6-layer based model with fully connected layer as a classifier (6FC) in our case) yields 99.14% accuracy. This best-performing model is also compared with the state-of-art models such as VGG-16, InceptionV3, and MobileNetV2, trained using the transfer learning approach. After model comparison, we found our best model (6FC) outperformed the other models based on evaluated performance metrics with improvements as 3.94% gain in accuracy, 25.97% reduced parameters, and less training time (0.51hr) and testing time (20.5 sec) compared to VGG-16. Comparing precision, recall, and f1-score values are also found high (between 0.98 to 1) with our proposed model. The weighted gradient class activation map (Grad-CAM) technique generates a visualization of class predictions on the test dataset. The Grad-Cam visualization of results validates the prediction score attained by the proposed model.
Dharmendra Kumar Mahato1, Amit Pundir2, Geetika Jain Saxena3
Babasaheb Bhimrao Ambedkar Bihar University, India1, Maharaja Agarsen College, India2,3
Convolutional Neural Network, Grad-CAM, Deep Learning, Data Augmentation, Transfer Learning
| January | February | March | April | May | June | July | August | September | October | November | December |
| 1 | 3 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| Published By :
ICTACT
Published In :
ICTACT Journal on Image and Video Processing
( Volume: 12 , Issue: 3 , Pages: 2615-2623 )
Date of Publication :
February 2022
Page Views :
126
Full Text Views :
11
|