Automatic Classification of Social Digital Media Images Using Convolutional Neural Networks

This study proposes an optimized CNN framework for social digital media image classification, addressing challenges in feature extraction and cross-platform adaptability. The model integrates ResNet50 with multi-scale convolution kernels <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(3\times 3,5\times 5,7\times 7)$</tex> to capture localized textures and global semantic features. Global average pooling (GAP) replaces fully connected layers, reducing parameters by 38%, while data augmentation and batch normalization (BN) enhance generalization. He initialization and adaptive Adam optimization stabilize training. Evaluated on a combined ImageNet-FFHQ dataset (20,000 images), the model achieves 91.43% accuracy and 89.7% F1-score, surpassing ResNet50 (89.81 %) and VGG16 (88.47%). It demonstrates efficient inference (21.7 ms/image) with reduced VRAM usage (6.4GB vs. ResNet50’ s 7.1GB), though its file size (127.5MB) limits lightweight deployment. The multi-scale design and GAP significantly improve feature robustness and computational efficiency. Future work will employ neural architecture search (NAS) to optimize model compression, enabling scalable deployment across resource-constrained platforms without sacrificing accuracy.