Automatic waste classification using deep learning and computer vision techniques

Abstract

Waste management refers to a system that starts with classifying different kinds of waste and gradually managing it from its inception to its final disposal. Labeling waste in a proper manner can ensure the best outcome of recycling. The reason we think that our thesis topic will bring about a positive change in the waste management system is because we are emphasizing on making the environment pollution free and reuse the waste as much as we can by classifying and detecting the recyclable stuff from the waste that are considered useless. A custom CNN model has been implemented in our paper to classify things more accurately. Here, we have utilized a large dataset “garbage classification” [19] with a big number of images but to train our model, we have used 8 different classes: battery, biological, cardboard, clothes, green-glass, paper, plastic, trash which have been augmented in order to make all the classes equal in size which has resulted in a total of 16,000 images.Pre-trained CNN models such as VGGNet16, Resent50, MobileNetV2, InceptionV3,EfficientNetB0 along with custom CNN models have been used and successfully achieved 87.57 percent, 94.34 percent, 96.99 percent, 95.71 percent, 35.92 percent,97.16 percent train accuracy and 89.38 percent, 94.34 percent, 96.81 percent, 94.47 percent, 36.75 percent and 97.58 percent validation accuracy respectively.Later on, the paper also evaluates the custom CNN model’s performance on an unseen test dataset via confusion matrix. In this study, we have also proposed YOLOv4 and YOLOv4-tiny with Darknet-53 as a method for the detection of waste. Here we have used the same dataset which we have used in the custom CNN model. During the testing phase, every model makes use of three different types of inputs, including videos, webcams and images.The outcome demonstrates that YOLOv4 exceeds YOLOv4- tiny in terms of object detection, despite YOLOv4-tiny’s advantages in aspects of computational speed.The best YOLOv4 results are mAP 85.73 percent, precision 0.78, recall 0.84, F1-score 0.81, and Average IoU 62.05 percent.The best YOLOv4- tiny results are mAP 81.28 percent, precision 0.60, recall 0.87, F1-score 0.71, and Average IoU 45.67 percent.