A mobile application to identify poison ivy (Toxicodendron radicans) plants in real time using convolutional neural network

Poison ivy (Toxicodendron radicans) is an important noxious plants in many countries around the world. Currently, the primary method used to identify this plant is based on visual evaluations of the leaf shape and arrangement. In this study, we evaluated four variations of YOLO-Tiny to identify Poison ivy and we proposed a smartphone application (APP) to perform the detection in real-time. Images were taken in different parks located in the United States and then augmented, totaling 3,407 images and 73,824 annotations. The model with the highest mean average precision (mAP), precision (P), recall (R), F1-score, and lowest loss function value was selected to build the APP. Two smartphones (Motorola M51 and Xiaomi Redmi Note 11 Pro) and five input resolutions were evaluated (224, 288, 320, 384, and 416 width and height pixels). The best performance during the training was achieved using a YOLOv4-Tiny-3 L architecture with mAP of 78.8%, 0.85 (P), 0.60 (R), 0.7 (F1-score), and 3.55 for the loss function. The APP evaluation revealed that at 224 and 288 pixels, F1-scores were lower (around 0.39) and fewer plants were identified but the FPS was relatively high (around 4.87). At 416 pixels, more plants were identified with high F1-scores (around 0.66), but with a low FPS rate (around 1.96). The best balance between performance and accuracy was observed at 340 and 320 pixels for both devices. Overall, the results suggest that YOLOv4-tiny-Tiny-3 L can successfully be deployed in smartphones to identify Poison ivy.