Toward sustainable clean ports: Smoke opacity measurement based on semantic segmentation and edge device

To overcome the limitations of manually determining the opacity of smoke emitted from ships, we present a novel framework using artificial intelligence and an edge device. Four deep learning models trained with black and white smoke images for semantic segmentation were compared and analyzed. It was confirmed that two of these models, namely the novel W-Net and DeepLab v3+m exhibited superior performance compared with the conventional models. DeepLab v3+m achieved the best performance as it improved localization performance by adding features through basic convolution layers and various transfer-learned models to the encoder and decoder, respectively. DeepLab v3+m was converted to an open neural network exchange model and used for the edge device. Two algorithms, the first of which saved images using a camera connected to the edge device, and the second, which generated a smoke opacity measurement report and transmitted the same to the smoke observer via e-mail, were executed simultaneously. Reports that accurately predicted and summarized the smoke images were delivered via e-mail at random times every hour. Finally, we proposed a policy framework for implementing the developed system in the field using smoke observers. A novel proposed policy framework based on DeepLab v3+m was demonstrated to be able to seamlessly and automatically detect smoke opacity from ships within a port.