SMPTS: Segmentation Method for Physically Touching Soybean Images

EasyChair Preprint 13071, version 1

Abstract

The study mainly aims to develop an image segmentation algorithm named SMPTS for the segmentation of physically touching soybean images in seed testing machines. SMPTS is a classical image-based method. In SMPTS, the binary method with an adaptive mean threshold is to divide the contours of soybeans. The medium filter eliminates salt and pepper noises on the binary image of soybean, which leads to less running time for SMPTS. Otsu with a fixed threshold is to extract the regions of interest of soybean images in physical touching. The minimum bounding rectangle locates individual seeds on the binary image. Individual seed images are cut from the physically contacted soybean images based on location and size. As a result, SMPTS could achieve more than 99% segmentation accuracy, with about 53ms for segmenting a soybean seed on NVIDIA Jetson TX2. Meanwhile, the segmentation accuracy of intact soybeans, immature soybeans, skin-damaged soybeans, spotted soybeans, and broken soybeans is about 100%, 99.66%, 96.56%, 99.44%, and 99.51%, respectively. The code for SMPTS is https://codeocean.com/capsule/9219546/tree/v1 under the MIT license.

Keyphrases: image processing, image segmentation, physically touching seeds, soybean seeds

@booklet{EasyChair:13071,
  author    = {Wei Lin and Jiarui Fang and Qin Su and Hongjian Liao and Shuo Liu and Heyang Yao and Peiquan Xu},
  title     = {SMPTS: Segmentation Method for Physically Touching Soybean Images},
  howpublished = {EasyChair Preprint 13071},
  year      = {EasyChair, 2024}}