Deep Learning-Based Binocular Image Analysis for In Situ Measurement of Particle Length Distribution During Crystallization Process

To resolve the challenge of measuring particle length distribution (PLD) during crystallization process via image analysis, a novel in situ measurement method with deep learning-based binocular image analysis is proposed in this paper. Firstly, a non-invasive binocular telecentric imaging scheme is designed for in situ measurement. Then a matching algorithm is presented to find the spatial correspondences between crystal images from a binocular image pair, by developing a telecentric stereo image rectification strategy together with a deep learning-based oriented object detection algorithm. Subsequently, the three-dimensional (3-D) space length of each identified crystal is reconstructed by integrating 2-D projected information from two matched crystal boxes in the binocular image pair, so as to acquire measurement accuracy on each crystal length. The proposed binocular imaging scheme and 3-D reconstruction method are validated by in situ measurement of a micro-checkerboard plate inserted into a cooling crystallizer. Experimental results on the cooling crystallization process of β-form L-glutamic acid demonstrate that the measurement accuracy on PLD is evidently improved with the mean absolute percentage errors smaller than 4.2%, compared to the recently developed methods.