RESUMO
This paper introduces a noise augmentation technique designed to enhance the robustness of state-of-the-art (SOTA) deep learning models against degraded image quality, a common challenge in long-term recording systems. Our method, demonstrated through the classification of digital holographic images, utilizes a novel approach to synthesize and apply random colored noise, addressing the typically encountered correlated noise patterns in such images. Empirical results show that our technique not only maintains classification accuracy in high-quality images but also significantly improves it when given noisy inputs without increasing the training time. This advancement demonstrates the potential of our approach for augmenting data for deep learning models to perform effectively in production under varied and suboptimal conditions.
RESUMO
A hologram, measured by using appropriate coherent illumination, records all substantial volumetric information of the measured sample. It is encoded in its interference patterns and, from these, the image of the sample objects can be reconstructed in different depths by using standard techniques of digital holography. We claim that a 2D convolutional network (CNN) cannot be efficient in decoding this volumetric information spread across the whole image as it inherently operates on local spatial features. Therefore, we propose a method, where we extract the volumetric information of the hologram by mapping it to a volume-using a standard wavefield propagation algorithm-and then feed it to a 3D-CNN-based architecture. We apply this method to a challenging real-life classification problem and compare its performance with an equivalent 2D-CNN counterpart. Furthermore, we inspect the robustness of the methods to slightly defocused inputs and find that the 3D method is inherently more robust in such cases. Additionally, we introduce a hologram-specific augmentation technique, called hologram defocus augmentation, that improves the performance of both methods for slightly defocused inputs. The proposed 3D-model outperforms the standard 2D method in classification accuracy both for in-focus and defocused input samples. Our results confirm and support our fundamental hypothesis that a 2D-CNN-based architecture is limited in the extraction of volumetric information globally encoded in the reconstructed hologram image.
Assuntos
Holografia , Imageamento Tridimensional , Redes Neurais de Computação , Holografia/classificaçãoRESUMO
We adopted an unpaired neural network training technique, namely CycleGAN, to generate bright-field microscope-like images from hologram reconstructions. The motivation for unpaired training in microscope applications is that the construction of paired/parallel datasets is cumbersome or sometimes not even feasible, for example, lensless or flow-through holographic measuring setups. Our results show that the proposed method is applicable in these cases and provides comparable results to the paired training. Furthermore, it has some favorable properties even though its metric scores are lower. The CycleGAN training results in sharper and-from this point of view-more realistic object reconstructions compared to the baseline paired setting. Finally, we show that a lower metric score of the unpaired training does not necessarily imply a worse image generation but a correct object synthesis, yet with a different focal representation.