In-situ Images Based Detection Methods for Crystallization Processes
Crystallization technology has been widely used for product separation and purification in chemical and pharmaceutical engineering applications.To facilitate production quality detection of crystallization process,in-situ image detection technologies have been concerned and developed in the recent years,gradually applied to real-time detection of crystal shape and size along with process optimization.However,the existing domestic and oversea methods of in-situ image detection are still in the initial application stage.The measurement accuracy and reliability are not high for the important indices of crystallization quality evaluation,such as crystal shape,size distribution and agglomeration degree during crystallization,while considerable long time is spent for crystal shape recognition and size measurement.Hence,these methods are difficult to be applied to crystallization process detection.Based on single-vision or binocular images captured in real time,this dissertation studies crystal shape recognition,two-dimensional(2D)and three-dimensional(3D)size measurement,and agglomeration degree detection during crystallization etc.The main research contents include:(1)Based on in-situ images captured in real time,an image analysis method is proposed for rapid recognition of crystal morphology.The method includes image preprocessing,feature analysis,effective sieving of crystal images,crystal shape classification.A shape feature,named Inner Distance Descriptor(IDD),is defined to classify different crystal shapes accurately,which does not depend on the crystal size and geometry direction.To improve the efficiency of real-time compution,high-dimensional feature vectors are reduced in dimension based on the Spectral Regression Kernel Discriminant Analysis(SRKDA),which can be effectively used to deal with the nonlinear problem of feature variables.Based on the Support Vector Machine(SVM),a crystal morphology classification method is established to ensure recognition accuracy.By using a non-invasive imaging system for an L-glutamic acid(LGA)cooling crystallization process,the results manifest that the proposed image analysis method can effectively detect two different polymorphic forms(prismatic a-form and needle-likeβ-form)during the crystallization process.(2)To facilitate on-line measurement of crystal size distribution during crystallization processes,a crystal 2D size measurement method is proposed based on sparse representation.A pixel equivalent calibration method based on subpixel edge detection and circular fitting is adopted to ensure the accuracy of size measurement.To overcome the adverse effects of uneven illumination and crystal motion caused by stirring the solution on in-situ captured images,a fast non-blind restoration method is adopted to improve the image quality.Meanwhile,a sparse representation dictionary is introduced to eliminate image noise.The structure edge information of crystal image is used as a prior knowledge to recover the images of crystals.The crystal image dictionary and fuzzy kernel are timely updated in terms of the crystal image quality parameters evaluated during a crystallization process,in order to enhance the restoration efficiency.Then the optimal fitting rectangle for segmented images is used to measure the crystal 2D sizes.The image measurement experiments of LGA cooling crystallization process are performed in comparison with off-line measurement of the sampled crystals with the electron microscopy,effectively demonstrating the accuracy and reliability of the proposed method for measurement of crystal size distribution.(3)For the crystallization process of LGA,a measurement method of crystal 3D size is proposed based on a binocular microscopic vision system.Two cameras are set up respectively at two different angles outside the crystallizer to capture images synchronously in real time.Correspondingly,each pair of binocular images is used together for reconstructing 3D crystal shapes.An image analysis method for determining crystal shapes is established,including image preprocessing,corner detection and corner matching.According to two different forms of LGA crystals(i.e.a-form and β-form),two different corner detection algorithms are proposed to identify the key corners,respectively.The corresponding 3D geometry model is established to approximately reconstruct the 3D shape of each identified crystal,and therefore the 3D sizes of each crystal are evaluated quantitatively.The effectiveness and practicability of the proposed method for 3D size measurement are verified by binocular image measurement experiments on the LGA crystallization process.(4)To detect crystal agglomeration degree that affects the growth quality during a crystallization process,a detection method of crystal agglomeration degree is proposed based on binocular images.The binocular images captured synchronously are segmented by an image preprocessing method.The IDD indices for different crystal shapes are defined to preliminarily sieve out crystal image regions including possible agglomeration.Then two texture descriptors are introduced to extract the features of image interest points and match the features of binocular images,so as to identify pseudo agglomeration.In addition,a fast algorithm for counting the number of crystals is given,which can calculate the number of unagglomerated crystals and the number of primary crystals in agglomerates.The results of KDP crystallization experiment demonstrate that the proposed method could accurately identify pseudo agglomeration of crystals and estimate the crystal agglomeration degree,while counting the number of unagglomerated crystals and primary crystals in agglomerates.