Y-Net: Insect Counting and Segmentation using Deep Learning on Embedded Devices

Insect pests can pose a serious threat to food production and agriculture in general and can cause substantial crop damage and economic losses. Monitoring insect pest populations is essential to control and mitigate these losses. Traditional monitoring methods are considered by growers and agronomists to be time-costly as well as labour-intensive tasks, which ultimately means that in times of high activity on farms it is a task which often is neglected. This study proposes an automated vision-based insect segmentation and counting approach through the use of deep learning (DL) models developed particularly for embedded systems. An image dataset for our target insect, Halyomorpha halys, was first created using images captured by our IoT-enabled image capture system deployed in a fruit orchard. Then, a Y-Net model inspired by U-Net was developed with the capability of insect counting in addition to segmentation. The performance of this model was assessed using a variety of different metrics, and the results demonstrated the feasibility and effectiveness of the model in counting and segmentation of insects using Edge-AI algorithms capable of running on embedded systems. Based on the achieved results, the proposed Y-Net model achieved a Mean Squared Error (MSE) of 1.9 for the insect counting task, an Intersection over Union (IoU) of 84.5% and a Dice Similarity Coefficient (DSC) of 91.5% for the segmentation task, with an inference time of nearly 0.4 seconds on a smartphone.