Performance Evaluation of Deep Learning Models for Real-Time Trash Classification on Embedded Devices

Trash classification represents a significant challenge in waste recycling industry, leading often to contamination due to manual misclassification of recyclable and non-recyclable items. This arises because human judgment relies on experience and knowledge rather than on objective criteria. To address this issue many scholars have proposed automated classification systems based on machine learning and deep learning (DL). All these works have not been tested on real time embedded systems. This paper proposes a comparison between several DL models to categorize waste into cardboard, plastic, glass, metal, paper, and general trash. Several DL architectures were retrained and implemented on a Raspberry Pi, i.e., MobileNet, ResNet, EfficientNet, and DenseNet models. The performance, i.e., classification accuracy, inference time and maximum number of frames per second (FPS) of these models were evaluated using the TrashNet dataset [4]. The obtained results demonstrate that EfficientNet B3 achieved the best performance based on compared parameters as well as a testing accuracy at about $86 \%$, while DenseNet achieved the lowest performance parameters with a low accuracy of $79 \%$. Additionally, the models were evaluated for their performance on embedded devices. According to the experiment conducted in this work, MobileNetV2 was the fastest at classification, while EfficientNet B7 was the slowest. Considering both accuracy and speed, EfficientNet B3 is the best model for real time applications.