An advanced deep learning pipeline designed to classify Indian monuments into eight distinct architectural styles. The project implements a multi-stage approach featuring Saliency-based localized cropping, Two-phase Transfer Learning, and Mixed-Precision training to achieve high accuracy on complex structural imagery.

• Saliency Processing: Utilizes Static Saliency Spectral Residual detection to identify and crop structurally significant features of monuments.
• Hybrid Dataset: Evaluates performance on both original images and a "Combined" dataset (Original + Saliency Crops).
• Dual Architecture Analysis: Comparative study between Inception V3 and Inception ResNet V2.
• Memory Optimized: Aggressive memory management (garbage collection, image resizing, and reduced batch sizes) to support 16GB RAM environments.

The dataset is categorized into 8 architectural classes. The following table provides the distribution of images across Training, Validation, and Test sets (based on the original image distribution).

Note: Saliency augmentation added an additional 6,002 localized crops to the "Combined" dataset experiments.

The pipeline employs a Two-Phase Transfer Learning strategy:

1. Phase 1: Feature Extraction: Base model frozen; training of custom Dense layers (512 units, Dropout 0.4) using a learning rate.
2. Phase 2: Fine-Tuning: Top half of the base model unfrozen; training with a learning rate for structural adaptation.

Comparative performance across four distinct configurations:

Best Performing Model: Inception ResNet V2 with Saliency Augmentation (96.73% Test Accuracy).

Summary of per-class performance for the best model (Inception ResNet V2 - Combined).

• Python: 3.9.23
• TensorFlow: 2.10.1
• Hardware: NVIDIA GeForce RTX 3050 Laptop GPU (Mixed Precision Enabled)
• Key Libraries: OpenCV (Saliency Module), TensorFlow Probability, PIL, Seaborn.

This research identifies that saliency-based localized cropping consistently improves classification accuracy by an average of 1.6% across architectures.