Lantana Camara Flower Detection Using an Improved Lightweight Convolutional Neural Networks in YOLOv5

Abstract

Detecting the presence of lantana camara flowers in the natural ecosystem is complicated as it requires rapid and precise flower detection. Existing detection models have a low detection rate, a sizeable false-negative rate, and a lack of algorithmic resilience and generalizability. This paper presents a detection model based on the enhanced You Only Look Once version 5 (YOLOv5) algorithm. First, a selective number of lantana camara flower images are gathered from publicly available dataset from iNaturalist, which contains more than 230,000 image from 5,000 different plant species to generate a dataset. Different loss functions were applied with YOLOv5m as the baseline. Then, images are randomly cropped, scaled, and arranged using the mosaic data augmentation process to generate new images. A mechanism for dynamic anchor boxes is proposed to address the erroneous anchor box prior information in YOLOv5. The dynamic anchor box module is introduced to the model, and the network training process dynamically modified the anchor box's size and position. By incorporating channel and spatial attention into the network's original structure, a novel attention mechanism is suggested to increase network detection performance. The detection performance of the enhanced method presented in this study is more than 4 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> greater than the mean average precision (mAP) of the baseline model, with the mAP score reaching 91.9% and the detection speed reaching 87 frames per second.