An Optimized CNN-SGD Framework for Multi-Label Image Classification in Online Educational Platforms

Abstract

The recent rise of online learning sites has led to a substantial increase in the amount of visual learning material in terms of diagrams, mathematical formulas, pages of text, and learning headers. These educational images need to be better structured and automatically tagged to enhance learning and provide better content access and tailoring. Nevertheless, single label classification methods cannot determine the multidimensional and overlapping characteristics of learning material, such as when a single image can belong to two or more categories at the same time. In order to overcome this drawback, this paper suggests an optimized multi-label image classification system that utilizes Convolutional Neural Networks (CNNs) with the Stochastic Gradient Descent (SGD) optimizer. To achieve multiple labels per image prediction as proposed, early activation with sigmoid functions in the output layer and binary cross-entropy loss are used to facilitate an independent prediction. The dataset is a collection of educational images separated into classes (including Maths, Text, Header and Diagram), and split into two subsets (80 percent training and 20 percent testing). Exploratory outcomes establish that CNN-SGD model attains classification with over 96% accuracy, high precision, recall and F1-score measure. This system has an easy to use web interface that fetishes real time image upload and auto multi label prediction. The suggested framework will provide scalable, efficient, and binding framework to intelligent educational content management within contemporary digital learning strategies.