Analisis Komparatif Arsitektur Deep Learning Untuk Aplikasi Computer Vision: Studi Literature Review

Abstrak

The development of computer vision technology has undergone a significant transformation with the emergence of increasingly sophisticated deep learning architectures. This study aims to conduct a comparative analysis of the characteristics, performance, and computational efficiency of seven prominent Convolutional Neural Network (CNN) architectures: LeNet-5, AlexNet, VGG, GoogLeNet, ResNet, SqueezeNet, and MobileNet, within the scope of modern computer vision applications. A systematic literature review was employed as the research methodology, analyzing scientific publications published between 2021 and 2025 from reputable databases such as IEEE Xplore, ScienceDirect, SpringerLink, and Google Scholar. The findings reveal that each architecture possesses unique strengths and trade-offs. LeNet-5 is effective for simple tasks; AlexNet introduced innovations such as the ReLU activation function and dropout regularization; VGG is notable for its network depth; GoogLeNet achieves efficiency through its Inception modules; ResNet addresses the vanishing gradient problem using skip connections; while SqueezeNet and MobileNet are optimized for mobile applications with limited computational resources. The study concludes that no single architecture is universally superior. Instead, optimal model selection depends on balancing accuracy, computational efficiency, and the specific resource constraints of the intended application.