DMDPD: deep hybrid model for debris plastic detection and classification with metaheuristic training

Abstract

Aquatic ecosystems are increasingly threatened by debris, which poses risks to aquatic life. Deep learning has emerged as a powerful tool for automated detection of debris. However, conventional methods often struggle to differentiate debris having similar shapes and textures. To address these challenges, this study proposes a deep hybrid model for debris plastic detection and classification (DMDPD) framework for accurate plastic debris classification in marine and aquatic environments. Initially, augmentation methods including contrast, brightness, saturation and sharpness are applied; and a bilateral filtering is then used for image preprocessing. Subsequently, improved mask R-CNN (Im.MRCNN) on a preprocessed image was applied to refine object boundaries. Additionally, improved significant local binary patterns (ImSLBP), gray level co-occurrence matrix (GLCM), modified binary patterns (MBP) and statistical descriptors were extracted from segmented regions. A hybrid model is proposed that combines deep belief network (DBN) and long short-term memory (LSTM) frameworks for classification. Moreover, a self-improved feedback artificial tree (SiFAT) algorithm is applied for adaptive parameter optimisation of both models. Experimental evaluation demonstrates that the proposed DMDPD framework outperforms traditional approaches, with an accuracy of 0.957 and a specificity of 0.972. Overall, the proposed method provides an effective solution for accurate aquatic and marine plastic debris classification, supporting improved strategies for mitigating aquatic and marine pollution.