Trash Classifier on RaspberryPI

A component that can classify trash items into their category. Uses Python and runs on a raspberryPI system.

Hardware Component parts

RaspberryPI:

The system was tested on RaspberryPI 3 model B.
The operating system is RaspberryPI OS Debian. 11 Bullseye.

Camera:

The camera used is ArduCam OV5647 Mini Camera Module.
It has a 5MP resolution.
We take still pictures of 2592x1944 resolution

LCD screen:

We used MHS3.5 LCD screen.
To output to the screen, we needed the driver found here: https://github.com/goodtft/LCD-show

Software Modules

app.py

The main application entry points.
Initializes the interface, the model, the camera, the hardware parts.
Allows all classes to interact together.

camera.py

The camera has multiple functionalities including starting a preview, stopping the preview, taking a photo

hardware.py

This module is responsible for the 4 leds, blue for general , red for metal, yellow for paper, green for plastic

interface.py

We used tkinter library to display the three different gui layouts, which are:

• The first layout: this layout will display bunch of text at the top, the camera frame, and the two buttons.
• The loading layout: this layout will display an animation which will present the loading. It keeps loading until the model finishes its classification.
• The final layout: this layout will display the classfication result after the item has been classified by the model. It will display the name of the category and the percentage.

model.py

Model Training

The model is trained on several public datasets of trash from Kaggle including:

1. https://www.kaggle.com/datasets/naidusurajvardhan/recycling-waste
2. https://www.kaggle.com/datasets/sapal6/waste-classification-data-v2
3. https://www.kaggle.com/datasets/imrshu/solidwaste
4. https://www.kaggle.com/datasets/hseyinsaidkoca/recyclable-solid-waste-dataset-on-5-background-co

Our dataset consist of images in 6691 total divided into: 4685 for training 959 for testing 1047 for validation

For the model we used VVG19 as a backbone for our classification The model can classify the images into one of 3 classes plastic, paper, and metal. Anything other than that with an accuracy less than 70% is classified as general.

Model class

Model(file_path) Constructor that takes the file path of the model and load it.

load_model_(model_filepath) takes the file path of the model and load it.

predict_img_file(img_filepath, enable_saving = True, display_result = False) takes the file path of the image, perform postprocessing, predict the class, and print the predictions.

postprocess_preds_(preds, display_result=display_result) print the list of predictions

send_imgs_and_preds() send the prediction results to the Raspberry Pi.

Requirements of model class:

TFLite version 2.10.0 or above

numpy version 1.23.5 or above

Pillow version 9.3.0 or above

Python version 3.9.2 or above

Requirements:

• tkinter
• rpi.gpio
• picamera
• tflite