Deep Learning Model for Segmentation of Okra Plants with Unconstrained Backgrounds

In this research, it is inclined to extract the Okra plant region from an unconstrained background using deep semantic segmentation network. A static thresholding model based on Hue, Saturation, and Value (HSV) space is used to create annotated dataset of images. A deep semantic segmentation module, U-Net, is trained and tested with camera-captured images of Okra plants with complex backgrounds to extract plant regions. The datasets of low to high-resolution images captured using smartphones which comprise a variety of early-stage bacterial wilt diseases of Okra plants are considered for experimentation The effectiveness of the segmentation module is assessed using metrics such as Intersection over Union (IOU), Dice coefficient, and Jaccard similarity index. Additionally, the segmented samples of images from U-Net are classified into healthy and disease-infected samples using the Inception V3 network. An IoU of 0.89, Jaccard index of 0.99, and Dice coefficient of more than 0.74 is obtained concerning segmentation efficiency.