Effective Scene Classification in Arabic Films Using Custom and Pre-trained CNN Architectures

Because of their rich visual diversity, unique cinematic styles, and the dearth of expert-annotated datasets devoted to Arab cinema, accurately classifying scenes in Arabic films is still a challenging task. A novel Convolutional Neural Network (CNN) architecture created particularly for Arabic film scene recognition will be discussed in the current research. The main goal of the study is to figure out how well the proposed CNN recognises and classifies Arabic scenes from films when compared with existing pre-trained architectures. According to the findings of six transfer learning models, a comparative analysis was carried out. We developed a new dataset that contains 28,298 images from ten different Arab films. The dataset is utilized to train the models for three epochs at the beginning. While training, we used the same hyperparameters, data splits, and preprocessing pipeline to train each model in order to ensure methodological consistency. Then, a comparison was made of the test performance of each model. Within 15 epochs, the created CNN had 97.23% test accuracy; thus, it showed high accuracy with comparatively little training effort. Among the evaluated pre-trained models on the AMSD dataset, the MobileNet model was the most successful with a 98% test accuracy. So the created CNN ranked second overall with a test accuracy of 97.23%, showing its usefulness for categorizing Arabic cinematic scenes. This paper highlights the importance of selecting appropriate model architectures and training strategies for scene classification, supported by quantitative ablation and confusion analyses.