Deep learning-based image processing framework for efficient surface litter detection in Computer Vision applications

Computer vision plays a vital role in automating environmental analysis by enabling real-time object detection and classification in diverse conditions. Litter pollution poses significant health and environmental risks due to inefficient disposal, and manual oversight is labor intensive. Effective litter detection is crucial for large-scale environmental monitoring. However, existing models face challenges such as the complexity of detecting shadowy objects in varying lighting conditions (e.g., during rain or under sun rays), difficulty in recognizing small objects, low accuracy, and poor real-time performance. Existing two-stage detectors, such as Faster R-CNN, also struggle with these issues. This paper introduces an automated deep learning-based image processing approach for accurate litter detection across different locations, using an enhanced version of YOLOv9s called LD-YOLOv9s. Key improvements in this novel approach include replacing convolutional layers with DynConvLayer in the backbone, integrating an SDConv-ADown module to substitute down-sampling layers in the neck, and using MPD-IoU instead of CIoU. These modifications reduce the chances of overlooking small objects, such as caps or lids, which had the least class meaning in the dataset, achieving a mAP of 78.3% with an inference time of 6.7ms. A significant contribution of this work is the LD-2024 dataset, curated from indoor and outdoor environments with manually annotated images. Performance comparisons were made with several YOLO versions (YOLOv3, YOLOv5, YOLOv6, YOLOv7, YOLOv8, YOLOv9 variants) and traditional object detectors (Faster R-CNN, Center Net, Retina Net, Cascade R-CNN). Ablation studies validate the effectiveness of LD-YOLOv9s, which outperforms conventional methods, achieving a 6.3% improvement in mean average precision (mAP) over YOLOv9s on the LD-2024 dataset.