Deep learning based processing framework for spatio-temporal analysis and classification of laser biospeckle data

• Published in: Optics and laser technology 2024-02, Vol.169, p.110138, Article 110138
• Main Authors: Thakur, Puneet Singh, Krejcar, Ondrej, Bhatia, Vimal, Prakash, Shashi
• Format: Article
• Language: English
• Subjects: Convolution neural network; Deep learning; Image processing; Laser biospeckle; Photonics; Seed infection
• ISSN: 0030-3992; 1879-2545
• DOI: 10.1016/j.optlastec.2023.110138

•3D convolution neural network is proposed for spatio-temporal analysis and classification of biospeckle data.•DL based numerical indexing strategy is developed for numerical quantification of biospeckle activity.•Impact of varying experimental parameters on accuracy of the proposed technique is analysed.•Model was tested and analysed on practical as well as computer generated synthetic data. Laser biospeckle is an advanced optical technique with the ability to non-destructively visualize various transient phenomenon via spatial and temporal statistics. However, accuracy of the existing image processing techniques used to process biospeckle data is hampered by various experimental and processing dependent factors. Therefore, in this work, a novel 3D convolution neural network (3D CNN) based deep learning (DL) architecture is developed for spatio-temporal analysis of biospeckle data in both qualitative and quantitative domains that effectively reduces errors introduced due to the influence of varying experimental parameters. Firstly, 3D CNN based image processing model is proposed for spatio-temporal analysis and classification of biospeckle data. Furthermore, a novel DL based numerical indexing strategy is developed for identification of level of activity in a sample. Finally, impact of varying experimental parameters on accuracy of the proposed technique is analyzed. In this direction, multiple experiments were performed to examine the effect of variation in input data parameters such as frame dimension, frame rate, number of frames, and background noise on accuracy of the trained model. Performance of the proposed model was analyzed and compared with respect to synthetic data generated by using rotating diffuser based simulation model. Robustness of the proposed strategy was also validated experimentally on practical data associated with identification of disease in seeds. Obtained results demonstrated that the proposed technique is accurate and can perform spatio-temporal classification and numerical indexing of the biospeckle data under varying experimental parameters.