I need an AI program that focuses on image recognition, specifically scenes, and turns each frame into actionable insights. The model should look beyond simple object detection and read the overall context—lighting, spatial relationships, and activity—to generate a concise summary or set of metrics that I can feed into downstream analytics. Scope • Curate or suggest a suitable scene-level dataset (open source or licensed). • Design and train a deep-learning model—CNN, transformer, or a hybrid approach—that excels at contextual scene understanding. • Provide an inference pipeline (Python preferred, PyTorch or TensorFlow) that accepts an image and outputs structured insights such as environment type, key activities, and confidence scores. • Include clear documentation: setup, training script, inference examples, and model performance benchmarks on a held-out test set. Acceptance • Top-1 scene context accuracy meets or exceeds published baselines for the chosen dataset. • Inference time per image ≤ 200 ms on a modern GPU. • Code is clean, reproducible, and packaged in a Git repository. If you have prior work in scene recognition, transfer learning, or vision transformers, I’d love to see it.