Energy Consumption Analysis of Deep Learning Model for the Recognition of Bengali Lexical Sign Language

In today's era, with the growing threat of global warming, it is essential to incorporate sustainability into deep learning models, specifically when it comes to carrying out heavy tasks such as the classification of images. In general, image data is heavier compared to numerical or even text data. Thus, it is essential to investigate the energy consumption and carbon footprint of these image classification models. To achieve this objective, this research focuses on building an energy-efficient deep learning model used to classify Bengali Lexical Signs. This dataset comprises 10,000 locally collected images that are trained using Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) as the primary model and other CNN pre-trained models, namely DensNet, MobileNet, VGG-16 and Inception V3. Before passing the images through the model, they are pre-processed to remove any noise and to induce robustness; data augmentation is applied. On the primary model, a k-4 cross-validation is also performed. The energy consumption data is collected during the training phase of each model using the Intel Power Gadget tool. From the energy consumption data, the carbon footprint is calculated using the value of the carbon intensity found in the UK for the prior thirty days. Further, a one-way ANOVA test, descriptive statistical studies, and the average power consumption of the processor are carried out using the energy consumption data to understand the behaviour of these models. According to the results, it can be deduced that the pre-trained models, namely Inception V3, VGG-16 and DenseNet, appear to have the lowest carbon footprint, while the primary model CNN-LSTM depicts a higher carbon footprint. Another interesting fact observed in this research is that Inception V3 and DenseNet models exhibit the highest energy consumption compared to that of other models. Therefore, from the study, it can be concluded that high energy is required for deep learning models which are used to process images and for other computer vision applications.