Waste Segregation Classifier using CNN (TrashNet Dataset)

Mixed waste is the quiet failure point of every recycling program. Bins get contaminated when a plastic bottle lands in the paper stream or a glass jar in plastic, and once mixed, whole batches get downcycled or landfilled because hand-sorting is slow, expensive and inconsistent. Campuses run awareness drives, but the moment of sorting — one person, one item, one bin — still depends on guesswork. This project turns that moment into a teachable computer-vision problem. It trains a convolutional neural network on TrashNet, the public 2,527-image dataset of waste photos collected by Gary Thung and Mindy Yang for Stanford's CS 229 course, to classify a photo of a waste item into six classes: glass, paper, cardboard, plastic, metal or trash. Transfer learning from MobileNetV2 keeps training cheap enough for a student machine while producing a model fast enough to demo live. The build includes the training notebook, evaluation with per-class precision/recall/F1 and a confusion matrix, and a browser demo that classifies uploaded or captured photos with confidence bars — the full pipeline from dataset to prediction, explainable in a viva.

1. TrashNet's 2,527 white-background photos (6 classes: glass 501, paper 594, cardboard 403, plastic 482, metal 410, trash 137) are split into training, validation and test sets with a stratified 80/10/10 split (design target); the small trash class gets class weighting to offset its imbalance.
2. Each image is resized to 224×224 px and normalized with ImageNet mean/std; during training, augmentation applies random rotations (design target ±20°), horizontal flips and zoom (design target ±15%) to make the model less sensitive to framing.
3. MobileNetV2 pretrained on ImageNet acts as the feature extractor; a classifier head (global average pooling → dropout → 6-way softmax) is trained on TrashNet with the backbone frozen.
4. After the head converges, the top convolutional blocks are unfrozen and fine-tuned at a lower learning rate (design target 1e-5), letting the network adapt its features to waste-specific textures.
5. The held-out test set produces per-class precision, recall and F1-score plus a confusion matrix — the notebook computes these from the actual training run; no accuracy is claimed before training.
6. For the demo, the trained weights are served (Flask inference API) and the web UI sends the uploaded or captured photo, receives the six softmax probabilities, and renders the top class with confidence bars. The standalone browser demo included for preview simulates inference for illustration.

Tech stack:

• Python · TensorFlow/Keras
• MobileNetV2 (ImageNet pretrained) · transfer learning
• TrashNet dataset · 2,527 images
• OpenCV · Pillow (preprocessing)
• NumPy · scikit-learn (metrics)
• Flask (inference API)
• HTML · CSS · JavaScript (demo UI)
• Jupyter notebook (training walkthrough)

Dataset & model details

• Dataset: TrashNet — public waste-photo dataset by Gary Thung & Mindy Yang (Stanford CS 229); 2,527 white-background photos (512×384 px) taken with iPhones under sunlight/room lighting. 6 classes with official counts: glass 501, paper 594, cardboard 403, plastic 482, metal 410, trash 137.
• Task: Single-label image classification — input a photo of one waste item, output one of 6 recyclability classes.
• Model: MobileNetV2 pretrained on ImageNet with a custom classifier head (global average pooling → dropout → 6-way softmax); two-phase training: frozen-backbone head training, then fine-tuning of the top convolutional blocks.
• Metrics: Per-class precision, recall and F1-score plus a confusion matrix on a stratified held-out test split, computed by the buyer's own training run; development target is F1 ≥ 0.80 on the test set. No accuracy is claimed before training.

• [Six-class waste classification] The trained CNN sorts a photo of a waste item into one of TrashNet's six classes — glass, paper, cardboard, plastic, metal or trash — giving a direct, visible answer to the "which bin?" question.
• [Transfer learning with MobileNetV2] The model starts from ImageNet-pretrained MobileNetV2; a new classifier head is trained on TrashNet first, then the top convolutional blocks are unfrozen for a short fine-tune so features adapt to waste textures.
• [Upload or camera capture] The demo accepts a JPG/PNG upload from disk or a webcam capture, resizes the photo to 224×224 and normalizes it with ImageNet statistics before inference, exactly as the training pipeline does.
• [Confidence bars across all six classes] Every prediction shows the softmax probability for each class, not just the winner — borderline cases like plastic-vs-glass make excellent viva discussion material.
• [Dataset facts panel] The demo carries the real TrashNet facts — dataset name, authors, source, 2,527 images and per-class counts — so the student cites the official repository, not guesswork.
• [Training walkthrough panel] Dataset split, augmentation, training settings and the evaluation procedure are laid out step by step, matching the training notebook the student actually runs.
• [Per-class evaluation metrics] The notebook reports precision, recall and F1-score per class plus a confusion matrix on a held-out test split; these numbers come from the buyer's own training run and are never promised in advance.

• Trained classifier weights + exported model file (SavedModel/ONNX)
• Training notebook: dataset loading, preprocessing, augmentation, transfer learning, fine-tuning and evaluation
• Inference demo web app (upload/camera, confidence bars, dataset and training panels)
• Flask inference API wiring for real predictions
• Project report PDF (dataset, methodology, confusion matrix and metrics from the training run)
• PPT presentation for final review
• Viva Q&A preparation document (CNN basics, transfer learning, class imbalance, evaluation metrics)