College of Computer Science and Mathematics

Discussion of a Doctorial Dissertation  in the College of Computer Science and Mathematics – Department of Mathematics Sciences entitled:

A hybrid model using nature inspired meta heuristic algorithms as a deep feature selection method

It was discussed in the discussion room at the Faculty of Computer Science and Mathematics at the  University of Mosul on Sunday  15-10-2023

Doctorial Dissertation by student Noor Muhammed Noori Abdulmuhsen Alnuaimi

, under the supervision of Prof. Dr. Omar Saber Qasim

                  

In recent years, significant efforts have been directed toward leveraging machine learning and deep learning methodologies to enhance image classification. The effectiveness of classification systems profoundly depends on the quality of features extracted from the images. This study delves into the realm of deep feature selection, aiming to maximize classification accuracy by identifying pivotal features that significantly influence the classification process. This endeavor is based on Convolutional Neural Networks (CNNs) in conjunction with meta-heuristic algorithms inspired by biological behaviors, such as the Marine Predator Algorithm (MPA), after its conversion from continuous to discrete space.

The research presents four distinct methodologies. The initial approach involves a hybrid fusion of the ResNet-18 CNN and the Binary Marine Predator Algorithm (BMPA). Subsequently, in the second methodology, the ResNet-50 CNN is combined with the BMPA. Seeking to refine the BMPA, the study integrates it with the Gray Wolf Optimization Algorithm (GWO) in a harmonious manner. This effectively combines equations from the second phase of BMPA with those of the GWO algorithm, resulting in a novel hybrid algorithm termed BMPA_GWO. The third approach manifests as a synergy between the ResNet-18 CNN and the BMPA_GWO algorithm, while the fourth strategy involves a similar fusion of the ResNet-50 CNN with the BMPA_GWO algorithm.

The experimentation phase involves the utilization of three diverse image datasets: Corn Leaf Diseases (NLB), Star-Galaxy Classification Data, and Is it Daisy? These datasets serve as the benchmark for evaluating the efficiency of the four proposed methodologies. This enables a comprehensive comparison of the third and fourth methodologies against the first and second approaches. Experimental results confirm the effectiveness of the BMPA_GWO hybrid algorithm in conjunction with convolutional neural networks. This validates its efficiency in improving image classification performance in terms of classification accuracy and the number of selected features.

 

The scientific committee included the following members:

prof.Dr. Ban Ahmed Hasan – Chairman

prof. Dr. Zakariya Yahya Algamal – Member

prof. Dr. Safwan Omar Hasson – Member

Ass. Prof. Zeyad Mohammed Abdullah – member

Ass. Prof.Dr. Mohammed Abdulrazaq Alkahya – Member

prof.Dr. Omar Saber Qasim – Member and supervisor