Detection and counting of pigment glands in cotton leaves using improved U-Net

Gossypol, as an important oil and raw material for feed, is mainly produced by cotton pigment gland, and has a wide range of applications in the fields of pharmaceutics, agriculture and industry. Accurate knowledge of the distribution of pigment gland in cotton leaves is important for estimating gossypol content. However, pigment glands are extremely small and densely distributed, manual counting is laborious and time-consuming, and difficult to count quickly and accurately. It is thus necessary to design a fast and accurate gland counting method. In this paper, the machine vision imaging technology is used to establish an image acquisition platform to obtain cotton leaf images, and a network structure is proposed based on deep learning, named as Interpolation-pooling net, to segment the pigment glands in the cotton leaf images. The network adopts the structure of first interpolation and then pooling, which is more conducive to the extraction of pigment gland features. The accuracy of segmentation of the model in cotton leaf image set is 96.7%, and the mIoU (Mean Intersection over Union), Recall, Precision and F1-score is 0.8181, 0.8004, 0.8004 and 0.8004 respectively. In addition, the number of pigment glands in cotton leaves of three different densities was measured. Compared with manual measurements, the square of the correlation coefficient (R ²) of the three density pigment glands reached 0.966, 0.942 and 0.91, respectively. The results show that the proposed semantic segmentation network based on deep learning has good performance in the detection and counting of cotton pigment glands, and has important value for evaluating the gossypol content of different cotton varieties. Compared with the traditional chemical reagent determination method, this method is safer and more economical.