Optimizing Litter Detection in Images: An Integration of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs)
Dhananjay Kumar, P. William, Shashikant Raghunathrao Deshmukh, A. Sharmila, Keshav Kaushik, J Shreyas
Abstract
Robust litter detection from photographs is important in realizing cleaner landscapes and efficient waste disposal. Current approaches typically cannot spot small or partly occluded trash in dense backgrounds such as cities and jungles. This research introduces a hybrid detection framework combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) via a Cross-Scale Adaptive Attention Fusion (CSAAF) process. While local space information in CNNs is extracted, global contextual knowledge in ViTs is also gained, and adaptive fusion between them increases accuracy and stability. Model assessments with mean Average Precision (mAP), Precision, Recall, and F1-score metrics result in enhanced detection of fine and hidden litter classes. In addition, the architecture keeps a light computational burden, guaranteeing real-time execution in diverse landscapes. This proposed solution demonstrates an efficient and scalable autonomous litter detection solution that enables intelligent environmental observation and waste disposal systems.