Abstract
•Deep learning-based photonics sensor is proposed for assessment of seed-quality.•The technique exploited biological features of seed for identification of quality.•Transfer learning-based processing framework is proposed for analysing speckle data.•Comparison of proposed sensor with the standard techniques was performed.•High accuracy (97.88%) was obtained for classifying high- and low-quality seeds.•The proposed technique is simple, non-destructive, fast, and automatic.
Seed-quality is one of the most important factors for achieving the objectives of uniform seedling establishment and high crop yield. In this work, we propose laser backscattering and deep transfer learning (TL) based photonics sensor for automatic identification and classification of high-quality seeds. The proposed sensor is based on capturing a single backscattered image of a seed sample and processing the acquired images by using deep learning (DL) based algorithms. Advantages of the proposed sensor include its ability to characterize morphological and biological changes related to seed-quality, lower memory requirement, robustness against external noise and vibration, easy alignments, and low complexity of acquisition and processing units. Furthermore, use of DL based processing frameworks including convolution neural network (CNN) and various TL models (VGG16, VGG19, InceptionV3, and ResNet50) extract abstract features from the images without any additional image processing and accelerate classification efficiency. Obtained results indicate that all the DL models performed significantly well with higher accuracy; however, InceptionV3 outperformed rest of the models with accuracy reaching up to 98.31%. To validate performance of the proposed sensor standard quality parameters comprising percentage imbibition (PI), radicle length, and germination percentage (GP) were also calculated. Significant change (p < 0.05) in these parameters show that the proposed sensor can accurately monitor the quality of seeds with higher accuracy. Moreover, experimental simplicity and DL based automatic classification make the sensor suitable for real-time applications.