Instance Segmentation for Local Rice Seed Germination Evaluation using Deep Learning

This study combines sophisticated deep learning technology to solve the manual procedure of rice seed germination evaluation which is time-consuming and error-prone. This paper examines the architectural behavior of CNN architectures in detecting, segmenting and classifying local rice seeds into normal, abnormal, fresh ungerminated and dead seeds. This is done using Convolutional Neural Network (CNN) architectures such as You Only Look Once (YOLO)v8s-Seg and YOLOv9c-Seg. A high-end smartphone was used to capture the image in natural light, and polygon annotation was used to label an image of seeds that covers individual, partially occluded and overlapping seeds in a single image. A post-processing technique called Segment Anything Model (SAM) was adopted to improve the mask borders of the rice seed objects for accuracy. The outcome illustrates that both models function well with a mean Average Precision (mAP@50) of more than 90%. However, there are misclassification cases identifying rice seeds between dead and fresh ungerminated, and then abnormal and normal. The YOLOv8s-Seg was shown to be more computationally efficient for mobile deployment where it has 40.2 Giga Floating Point Operations Per Second (GFLOPs). Meanwhile, YOLOv9c-Seg was found to be better for feature extraction and fast convergence. However, YOLOv9c-Seg was computationally complex for mobile edge deployment. The research reveals that the incorporation of automated seed quality assessments from an artificial intelligence (AI) framework to empower the process can drastically ramp up agricultural productivity and sustain economic growth.