HGB-CBSNet: hybrid gradient boosting based coot bird search network for solid waste classification in IoT-enabled smart waste management system

The internet of things (IoT) paradigm plays a significant role in improving smart city tracking applications and real-time city management. Solid waste management, which can have negative influence on health and environment of society, has been an important issue associated to smart city applications. Classifying the solid waste materials into biodegradable and non-biodegradable facilitates their conversion into useful energy and appropriate disposal. The progression of efficient computational methods such as image processing and artificial intelligence provides various solutions to address the identified management issues. However, the low accuracy, high computational complexity and high error rate of existing waste classification techniques for solid waste still remain unsolved. In this paper, a novel hybrid gradient boosting based coot bird search network (HGB-CBSNet) is proposed for effective classification of solid waste images. Initially, pre-processing is carried out using upgraded weighted mean filtering (UWMFil) and upgraded homomorphic filtering (UHFil) to remove unwanted noise and contrast enhancement. After pre-processing, a pyramid-dilated DenseNet model is applied to extract significant features from the pre-processed image. Further, the optimal features are selected using improved coot bird search algorithm (ICBSA) to minimize the computational complexity and redundant features. Finally, based on the selected features, the classification of solid wastes is performed through hybridized light gradient boosting machine and an extreme gradient boosting scheme (LGBM and XGBoost). The simulation of proposed HGB-CBSNet model is accomplished using python platform and the experimental results are verified in terms of dissimilar performance indicators through publically available TrashNet dataset. As a result, the HGB-CBSNet model has reached maximum accuracy of 99.72% with reduced complexity as compared with the existing methods.