The rapid increase in municipal solid waste driven by urbanisation, industrialisation and population growth has created significant environmental and public health challenges. Effective segregation is the first and most critical step in recycling, yet conventional sorting depends on manual labour and is slow, costly and inconsistent. Deep learning has enabled automated waste classification with high accuracy, but existing methods suffer from limited dataset diversity, weak generalisation under complex environmental conditions, high computational cost and inadequate real-time deployment capability, and most address classification alone without offering actionable recycling guidance. This paper presents the design and development of a hybrid deep learning framework for intelligent waste classification and recycling recommendation. The framework is organised into six stages and employs EfficientNetV2B0 as the feature extraction backbone, initialised with ImageNet weights and adapted by a two-phase transfer learning schedule in which the classification head is trained first and the upper backbone blocks are then fine-tuned at a reduced learning rate. Comprehensive pre-processing, geometric and photometric augmentation, global average pooling and a Softmax classifier support multi-class recognition of recyclable, organic and non-recyclable materials, and an integrated product recommendation module converts the predicted category into recycling, reuse and disposal guidance. Three algorithms specify the pipeline, the fine-tuning schedule and the recommendation module, and evaluation uses accuracy, precision, recall, F1-score, specificity, the Matthews correlation coefficient, Cohen’s kappa and AUC-ROC. The proposed framework attains 91.87% accuracy, 91.65% precision, 91.92% recall, an F1-score of 91.78% and 99.71% specificity, with an MCC of 0.903, a kappa of 0.901 and an AUC-ROC of 0.978, exceeding MobileNetV2 by 9.46 percentage points, ResNet50 by 7.74, DenseNet201 by 6.85, EfficientNetB0 by 5.33 and plain EfficientNetV2B0 transfer learning by 3.41. Class-wise analysis localises the residual error to the non-recyclable category, whose error rate of 11.47% is roughly double that of plastic, which identifies intra-class variation rather than backbone capacity as the binding constraint.

Sapna Pandey, Neelesh Shrivastava, "Design and Development of a Hybrid EfficientNetV2B0 Deep Learning Framework for Intelligent Waste Classification and Recycling Recommendation" International Journal of Scientific Research and Engineering Development, V9(5): Page(299-305) September - October 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.