A COMPREHENSIVE APPLICATION OF ARTIFICIAL INTELLIGENCE BASED HYBRID DEEP LEARNING IN IMAGE CLASSIFICATION
Loading...
![Thumbnail Image](assets/images/Item.jpg)
Date
2023-02
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
G. B. Pant University of Agriculture & Technology, Pantnagar-263145
Abstract
Artificial intelligence refers to the simulation of human intelligence in machines that are
programmed to think like humans and mimic their actions. The term may also be applied to any machine
that exhibits traits associated with a human mind such as learning and problem-solving. Deep learning
drives many artificial intelligence applications and services that improve automation, performing
analytical and physical tasks without human intervention. Real-world deep learning applications are a
part of our daily lives, but in most cases, they are so well-integrated into products and services that users
are unaware of the complex data processing that is taking place in the background. The applications for
artificial intelligence are endless. In light of the foregoing facts, the present research is centered to
propose a comprehensive application of hybrid deep learning models based on artificial intelligence
techniques which include transfer learning, deep learning, machine learning, and fuzzy inference system.
In this study, six models are introduced. Model [1] presents an effective facial emotion intensity
classifier by fusion of the pre-trained deep architecture and fuzzy inference system. The pre-trained
architecture VGG16 is used for basic emotion classification and it predicts emotion class with the class
index value. By class index value, images are sent to the corresponding Fuzzy inference system for
estimating the intensity level of detected emotion. This fusion model effectively identifies the facial
emotions (happy, sad, surprise, and angry) and also predict the 13 categories of emotion intensity. Model
[2] proposes a casting fault detector to automate the inspection process in casting production. two
transfer learning-based convolution neural networks and one lightly structured convolution neural
network are created and verified for detecting manufacturing flaws in submersible pump impellers.
Model [3] investigate various fusion of deep learning-based feature extractors and machine learningbased
classifiers for differentiating two subtypes Adenocarcinoma and Squamous cell carcinoma of nonsmall
cell lung cancer. This research generates seven automated integrated models by using five deep
feature extractors: InceptionResNetV2, InceptionV3, Xception, VGG16 and, VGG19 and three
classifiers: Support vector machine, XGBoost, and Fully connected neural network. Then the most
effective optimal model is selected via comparative study and performance. Model [4] construct three
transfer learning-based classifiers Binary, Benign, and Malignant. The Binary classifier classifies breast
cancer as benign and malignant, the Benign classifier classifies four sub-classes of benign cancer and the
Malignant classifier classifies four sub-classes of malignant cancer. All three classifiers are individually
trained for their corresponding classification task and then integrated to give the outcome of the
combined proposed system. As a result, the proposed system automatically classifies cancer into its
major class and then sub-class with greater accuracy. The proposed breast cancer classification is
performed on BreaKHis and BACH data. Model [5] develop an automated optimal novel model for knee
osteoporosis classification through a comprehensive examination held on various deep learning
architectures. This examination starts with the transfer learning model (ResNet50, MobileNet,
ResNet50V2, and Xception) and then extended to a concatenated version of the top two models
MobileNet and ResNet50V2 based on performance. Finally, on the basis of critical examination, an
optimal novel CNN inspired by MobileNet and ResNet50V2 is invented. For classification, FCNN
optimized by random search technique is employed. Model [6] develop the Cascade network to identify
diabetic retinopathy on the most challenging data IDRiD in literature. The Cascade network extracts the
multiscale dense features through the task specific novel CNN and Of-the-self CNN Xception. The
proposed model first time in literature construct an architecture to perform combinedly binary and multigrading
of DR and DME in a single input. The present study implements a limited data merging strategy
with the most relevant DDR data images. To find the optimal network, this research also investigates the
various network architectures and Cascade networks with different Of-the-self CNN.
Description
Keywords
Citation
Theses of Ph.D