Sep 2024/AI/ML/course
Animals10 Image Classification Project
Building a computer vision pipeline is often overcomplicated by deployment ambitions before the baseline is even proven. I built this CNN-based classifier to intentionally narrow the massive Animals10 dataset into three classes (chicken, dog, spider). The goal was to keep the learning workflow tightly constrained, understandable, and highly auditable.
I designed a notebook-driven TensorFlow and Keras workflow that enforces a strict 80/20 data split. The model successfully isolates features across the three classes and meets the 85 percent accuracy target. Instead of chasing a complex cloud deployment, the pipeline exports the trained network into multiple lightweight formats like TF Lite and TensorFlow.js for local inference.
The biggest takeaway is that reproducible pipelines matter far more than flashy deployments. By preserving the exact environment requirements alongside the model artifacts, anyone can instantly verify the training process. This disciplined approach ensures that the computer vision baseline is actually reliable before it ever touches a production server.