Book Cover Classification
Systems

Visual design in publishing relies on deeply ingrained genre signifiers—from the typographic layout of academic textbooks to the high-contrast color palettes of science fiction.

Rather than relying on pre-trained black-box models like ResNet or CLIP, this project focused on designing and implementing a Convolutional Neural Network (CNN) entirely from scratch using raw mathematics and NumPy. The objective was to understand the exact mechanics of spatial feature extraction, backpropagation through convolutional layers, and parameter optimization.

The dataset comprised over 50,000 book cover images across 10 primary categories. The pipeline actively analyzes visual hierarchy, color histograms, edge density, and text-to-image ratios to classify titles with minimal computational overhead.

Constructing the backward pass for 2D convolutions required deriving custom vectorized gradient calculations to prevent memory bottlenecks during training. The final architecture utilized modular residual blocks to combat vanishing gradients across deeper layers.

An interesting mathematical byproduct of the classification mapping was the discovery of distinct clustering in latent space between seemingly unrelated genres, such as historical fiction and high-fantasy, driven primarily by typographic serifs and warm sepia color grading.