Every component in Recycla was chosen for a reason. From the neural network architecture to the servo motors, here is the full technical breakdown of a system designed to classify waste in under 200 milliseconds on a $35 computer.
Pre-trained on 1.2 million images by Google, then fine-tuned on our custom waste dataset. The model classifies six material categories: cardboard, glass, metal, paper, plastic, and general trash.
MobileNetV2 was chosen for its depthwise separable convolutions, which reduce computation by 8-9x compared to standard convolutions while maintaining classification accuracy. This makes it ideal for inference on resource-constrained hardware.
4GB RAM, quad-core ARM Cortex-A72. Runs TFLite inference, controls GPIO for sensors and actuators, and manages the full classification pipeline. Chosen for its balance of processing power, GPIO accessibility, and cost-effectiveness.
Sony IMX219 sensor. Captures sharp images across varying lighting conditions for reliable classification of glossy, matte, and transparent surfaces. The CSI interface provides low-latency image transfer directly to the Pi's GPU.
Detects object presence within 30cm to trigger the classification pipeline. Low power consumption allows continuous standby monitoring. Dual transducer design provides reliable distance measurement with 3mm precision.
Independent servo control for recycling and garbage compartment lids. 180-degree rotation, 1.8kg/cm torque. Responsive enough for real-time sorting, reliable enough for thousands of cycles.
Our training pipeline uses aggressive data augmentation to ensure the model performs outside of clean lab conditions.
Random shifts simulate fluorescent, natural, and low-light conditions found across campus buildings.
Rotation, flips, and zoom account for varied item orientations as waste is dropped into the bin.
Channel shifts handle the difference between warm indoor lighting and cool outdoor daylight.
Specular highlight training addresses glossy packaging surfaces that can confuse standard classifiers.
The dataset is designed to grow. Adding new training data is as simple as dropping photos into the dataset folder and retraining.