:::

Establish a digital platform for seed illustration

This project aims to establish a digital platform for seed illustration (hereinafter called the “seed platform”) and an inspection system for rice seeds with image processing (hereinafter called the “rice system”). The purpose of the rice system was to establish a rapid and accuracy inspection system for rice seeds with image processing in order to maintain the quality and purity of rice seeds under seed propagation system, and decrease the time and manpower for seed purity analysis. In this year, the inspection system focused on rice seed image-database in 4 varieties including ’Tainan 16’、’Kaohsiung sen 7’、’Taichung sen 17’ and ’ Taichung sen wax 2’. Consequently, the image catching rate was equal to 91.5%, and the accurate rate of variety identification equaled to 95.4%. 

For seed platform, using image processing technology to 118 plant seeds in 35 families including Solanaceae, a multi-modal hierarchical classifier is used to identify plant seeds, and five different attributes of characteristic graphic information are captured, including texture, color, and shape in the image increase the diversity and complementarity of the data. The important information of the image is represented by low-dimensional vectors, and the self-organizing mapping neural network is used to correspond to its feature mapping coordinates, which will be used in a hierarchical manner. The amount of data is reduced, and the identification performance is output on the premise of the shortest time to achieve real-time inspection and application. The results show that when the final decision is made to merge PNN, 85% is the average identification result, and TOP-5 as the identification rule, the identification rate can be as high as 95%, the calculation time and resources used are relatively small, and the average of the samples are processing time is 55 milliseconds per sample.

The digital platform for seed illustration
▲Fig. 1. The digital platform for seed illustration
The deep learning classifier architecture for multimodal feature of seed
▲Fig. 2. The deep learning classifier architecture for multimodal feature of seed