Experimental particle’s shapes reconstructions from their interferometric images using the Error-Reduction algorithm

Particle Shape Recognition with Interferometric Particle Imaging Using a Convolutional Neural Network in Polar Coordinates

“…This is a very satisfying result for image classification. The addition the panel of non-centrosymmetric shapes was important because interferometric particle imaging can suffer from the twin image problem, as discussed using phase retrieval algorithms in a recent study [24]. All these patterns were experimental patterns obtained with the DMD setup.…”

“…However, they present limitations. For example, reconstructions using the error reduction algorithm have been shown to be less accurate in the case of non-centrosymmetric particles due to the twin image problem [24]. The reconstruction of a Y-shaped particle can lead to a particle with six branches by the superposition of a Y-shaped particle with its symmetric twin image.…”

“…It is a Thorlabs BC106N-VIS/M camera with 1360 × 1024 pixels and a pixel size of 6.45 µm × 6.45 µm. The parameters of the setup are those of reference [24].…”

“…In the case of rough particles, the interferometric image is a speckle pattern whose 2D Fourier transform is linked to the global shape of the particle [18]. The analysis of the speckle patterns can be limited to Fourier transforms to determine the size of a particle or to other algorithms, such as phase retrieval algorithms [22][23][24], to identify the shape. Deep learning offers a very interesting alternative [25][26][27], although it requires a very large amount of data for training, and such data are difficult to obtain.…”