Zwha Abdulhmid Mohamed Albeerish

Abstract:

This study, titled “Deep Learning-Based Image Classification Using MATLAB for Real-World Applications”, presents an integrated computer framework for developing and evaluating image classification systems using the MATLAB environment. The study aims to bridge the “semantic gap” resulting from the inadequacy of traditional computer vision methods based on manual feature extraction, which suffer from deterioration when applied in complex real-world environments.  To overcome the problem of the scarcity and difficulty of obtaining specific data in specialized sectors such as clinical medicine and precision agriculture, the study employs “Transfer/Learning” technology. The research is based on a quantitative experimental design to compare the performance and athletic behavior of four pre-trained convolutional neural structures: SqueezeNet, GoogLeNet, ResNet-50, and VGG-19. The results showed a clear trade-off between architecture efficiency and inferential accuracy: the VGG-19 achieved the highest overall accuracy of 100.0% but with a huge memory consumption (535.0 MB), while the SqueezeNet model was characterized by its lightness (4.6 MB) and inference speed (4.2 ms), making it ideal for peripherals. In contrast, the ResNet-50 model provided excellent balance with an accuracy of 98.5%, justifying its adoption in medical diagnostic tasks.  The study also looks at the dynamics of numerical enhancers: the Adam optimizer showed a fast and optimal numerical convergence for small groups, while the SGDM optimizer provided better generalization for complex groups. In terms of data processing, positional processes proved to be highly efficient, as the conversion of agricultural images to the HSV color space contributed to separating light values from colors and reducing the effect of shadows, while the Grayscale \ Replication mechanism enabled the adaptation of diagnostic medical images to meet the requirements of deep models. The study recommends selecting models based on hardware constraints, and moving towards segmentation to improve classification accuracy in complex real-world environments.