Approach for identifying crop seeds with similar appearances using hyperspectral images and improved ResNet 18 based on cloud platform

Abstract

Abstract This study proposes an improved ResNet18 model based on cloud platform and hyperspectral image to identify crop seeds. Hyperspectral images are preprocessed by moving average method (MA) and standard normal variable transformation (SNV) to reduce spectral data interference. The successive projection algorithm (SPA) is used to select characteristic wavelengths and reduce the band dimension. The heterogeneous convolution and channel attention mechanism (CAM) are used to reduce parameters and computation load of ResNet18. Image recognition with improved ResNet18 takes 4.4 s with 4.2 s improvement over the original ResNet18 and saves 2808 KB memory. Accuracy has increased from 93.1% to 96.7%. The improved ResNet18 was deployed on Alibaba Cloud ESC server with uWSGI, Nginx and flask to achieve high concurrency and fast recognition. For single‐variety seed identification, the average time reduced to 0.56 s, with an 87% decrease and accuracy is 96.3%. For mixed seeds of five types, accuracy is 90.22% with a time consumption of 0.88 s. The results demonstrated that the proposed method exhibited high effectiveness and accuracy.

Keywords

• clouds
• convolutional neural nets
• hyperspectral imaging