Optimization Enhancements for Faster R-CNN in Floating Bottle Detection

Trash bottles which are floating is one of the main causes of the worrying ecological threat posed by the developing pollution of inland water bodies. Real-time object detection algorithms in aggregation with unmanned boats can propose ways to eliminate floating trash from rivers. The study presents an extensive analysis of Faster RCNN(F-RCNN) models by applying different backbones, such as ResNet50, VGG16, DenseNet and mobileNetV3 and different optimizers such as Adam, Stochastic gradient descent for bottle identification in inland water. FloW_Img [1] is a dataset of floating trash in the river, and the performance metrics like mean average precision(mAP), recall and precision are used, to compare and experimentally validate each algorithm's detection efficiency and accuracy. The approach trains each architecture for 40 epochs. Regarding the precision of detection, the experimental results shows that F-RCNN with resnet50 as a backbone and Stochastic gradient descent (SGD) optimizer has high detection mean average precision (mAP) with mAP@[IoU=0.5] of 18%