Effect of Convolutional Depth on Image Recognition Performance: VGG vs. ResNet vs. GoogLeNet

Increasing convolutional depth has been a cornerstone of advances in image recognition, yet the relationship between a network’s layer count and its per formance remains non-uniform. This paper presents a controlled comparative study of three canonical architectures — VGG, ResNet, and GoogLeNet —to isolate the effects of depth on classification accuracy, convergence dynamics, and computational efficiency. By standardizing training protocols and introducing a formal mathematical distinction between nominal depth (Dnom) and effective depth (Deff), we demonstrate that the benefits of scaling depend critically on architectural mechanisms that stabilize the optimization landscape. Our results reveal that while plain sequential networks like VGG suffer from early accuracy saturation and gradient attenuation, residual- and Inception-based designs successfully leverage high nominal depth by maintaining a lower, more manageable effective depth. We provide empirical and theoretical evidence that Deff is the primary factor determining whether depth serves as a productive scaling dimension, offering a consistent metric for cross-architecture comparison.