Volume 90, Issue 7-8 - Special Issue: Image Processing Forum 2022/ Forum Bildverarbeitung 2022

Imaging through turbid media leads to a great loss of information decreasing the image quality. This loss of quality is due to the light interaction with the medium, since part of the light will pass through it, but another part will be scattered generating a blurred image pattern. In this work we try to decrease this problem by adding an absorbent, eliminating part of the scattered radiation responsible for the turbidity. With this aim, we use graphene nanoplates as an absorbent and we also compare the experimental results with black carbon powder.

Diffraction based multispectral imaging is a method for obtaining snapshot multispectral image information by using a structured diffraction grating or computer-generated hologram in an intermediate image plane of an imaging system. Without the need of complicated filter manufacturing it becomes possible to realize application specific spatio-spectral sampling of an image. The spectral resolution is coupled with the spatial resolution because the central stop of the imaging system serves also as the main spectral filtering component. By using two image sensors and the zeroth order of the computer-generated holograms it becomes possible to obtain high resolution spatial information as well as spectral information on a coarser sampling grid.

In many fields and applications a spatially resolved measurement of the refractive index is desirable. This also applies to the additive manufacturing of optical components and is especially true in the special case of photopolymerization of liquid resins by UV irradiation. The key point here is that the resulting refractive index of the optical component to be printed depends on the degree of cure of the polymer. The latter is directly related to the UV irradiance used for curing. Thus, a local variation of the irradiance can enable a local variation of the refractive index of the additively manufactured optical component. In order to set a defined refractive index distribution, it is desirable to measure the refractive index during the photopolymerization of the resin in a temporally and spatially resolved manner. In this article, a method is discussed with which this can be achieved. The basic principle is that the polymer under investigation is brought into contact with a prism and, based on the principle of total reflection at the interface between prism and polymer, the refractive index of the sample can be inferred.

Image-based object detection is a crucial task in autonomous driving. In many cases, objects are not correctly detected and classified if they are only partially visible due to a limited field of view. Also, even if stitched panoramic images are used, errors in object detection can still occur if the seam between individual images is visible. This happens due to vignetting or different exposure, although the images are optimally aligned. In this article, we present a real-time capable and effective method for vignetting compensation and exposure correction. Before runtime, the camera response function is determined and the vignetting model is preliminarily approximated. We obtain the irradiance from the intensity values of incoming images. Then, the vignetting model is applied. Afterwards, the pixels at the seam are used to correct the exposure. Finally, we convert the corrected irradiance back to intensity values. We evaluate our approach by measuring the image stitching accuracy in the overlapping area by the IoU of grayscale histograms and the mean absolute error of intensity values. The metrics are applied both on data recorded with our experimental vehicle and on the publicly available nuScenes dataset. Finally, we demonstrate that our approach runs in real-time on GPU.

The size and number of images and the amount of data we process every day have grown rapidly over the last years. Quantum computers promise to process this data more efficiently since classical images can be stored in quantum states. Experiments on quantum computer simulators prove the paradigms this promise is built on to be correct. However, currently, running the very same algorithms on a real quantum computer is often too error-prone to be of any practical use. We explore the current possibilities for image processing on real quantum computers. We redesign a commonly used quantum image encoding technique to reduce its susceptibility to errors. We show experimentally that the current size limit for images to be encoded on a quantum computer and subsequently retrieved with an error of at most 5 % is 2 × 2 pixels. A way to circumvent this limitation is to combine ideas of classical filtering with a quantum algorithm operating locally, only. We show the practicability of this strategy using the application example of edge detection. Our hybrid filtering scheme’s quantum part is an artificial neuron, working well on real quantum computers, too.

An almost lossless compression method for images is introduced adapted to the temporal noise of image sensors. In a first step, a non-linear gray value transform is applied to generate an image with a gray value independent temporal noise and less bits than the original image. The chosen value for the standard deviation of the temporal noise in the transformed image determines how accurately mean values and the standard deviation of temporal noise can be computed and to which extent the image can be compressed further by a lossless compression in a second step. Just a measurement of the noise characteristics according to the open and international EMVA standard 1288, a non-linear gray value transform for noise equalization, and an open source lossless compression algorithm are required to use this new compression method.

We have developed a novel metric to gauge the performance of artificial intelligence (AI) or machine learning (ML) algorithms, called the Spatial Recall Index (SRI). The novelty is the spatial resolution of a standard performance indicator, as a Recall value is assigned to each individual pixel. This generates a distribution of the performance of a given AI-algorithm with the resolution of the images in the dataset. While the mathematical basis has already been presented before, here we demonstrate the usage on more datasets and delve into in-depth application examples. We examine both the MS COCO and the Berkeley Deep Drive datasets, using a state-of-the-art object detection algorithm. The dataset is degraded using a physical-realistic lens-model, where the optical performance varies over the field of view, as a real camera would. This study highlights the usefulness of the SRI, as every image has been taken by realistic optics. A generalization, the GSRI is introduced, from which we derive SRIA, weighting with object area and SRIrisk intended for autonomous driving. Finally, these metrics are compared.

Anomaly detection with machine learning in industrial inspection systems for manufactured products relies on labelled data. This raises the question of how the labelling by humans should be conducted. Moreover, such a system will most likely always be imperfect and potentially need a human fall-back mechanism for ambiguous cases. We consider the case where we want to optimise the cost of the combined inspection process done by humans together with a pre-trained algorithm. This gives improved combined performance and increases the knowledge of the performance of the pre-trained model. We focus on so-called one-class classification problems which produce a continuous outlier score. After establishing some initial setup mechanisms ranging from using prior knowledge to calibrated models, we then define some cost model for machine inspection with a possible second inspection of the sample done by a human. Further, we discuss in this cost model how to select two optimal boundaries of the outlier score, where in between these two boundaries human inspection takes place. Finally, we frame this established knowledge into an applicable algorithm and conduct some experiments for the validity of the model.

Sensor-based sorting offers cutting-edge solutions for separating granular materials. The line-scanning sensors currently in use in such systems only produce a single observation of each object and no data on its movement. According to recent studies, using an area-scan camera has the potential to reduce both characterization and separation error in a sorting process. A predictive tracking approach based on Kalman filters makes it possible to estimate the followed paths and parametrize a unique motion model for each object using a multiobject tracking system. While earlier studies concentrated on physically-motivated motion models, it has been demonstrated that novel machine learning techniques produce predictions that are more accurate. In this paper, we describe the creation of a predictive tracking system based on neural networks. The new algorithm is applied to an experimental sorting system and to a numerical model of the sorter. Although the new approach does not yet fully reach the achieved sorting quality of the existing approaches, it allows the use of the general method without requiring expert knowledge or a fundamental understanding of the parameterization of the particle motion model.

This contribution proposes an approach and the respective tool that uses Active Deep Learning (ADL) to segment industrial three-dimensional computed tomography (3D CT) data. The general approach is application-independent and includes an iterative human-in-the-loop Active Learning (AL) process that produces labeled training data and a trained Deep Learning (DL) model for semantic segmentation. The model is continuously improved during iterations such that manual labeling effort is reduced. In addition, the user can minimize user interaction with the aid of a random forest-based classifier and focus on unclear or invalid segmentation results. The complete workflow is implemented within one single Python tool. The approach is demonstrated in detail for two industrial use cases: Single fiber analysis and plant segmentation. For plant segmentation, the method is compared to a baseline and a classic image processing algorithm.

The increasing use of automated laser welding processes causes high demands on quality control. 2D or 3D sensor technology can be used for data acquisition to monitor the weld quality after laser welding. Compared to a 2D camera image, the 3D height data, e.g. acquired using optical coherence tomography, contains additional relevant information for quality inspection. However, the disadvantages are system complexity, higher costs, and longer acquisition times. Therefore, we compare image-based methods with the quality assessment based on height data. The first method uses feature vectors from grayscale images taken coaxially with the laser beam. The significant advantage is that a camera is often integrated into the laser system, so no additional hardware is required. In the second approach, we use an AI-based single-view 3D reconstruction method. The height profile is reconstructed from a camera image and used for further quality assessment. Thus, we combine the advantages of 2D data acquisition with higher accuracy in evaluating 3D data. In addition, we consider the usually low data availability in the industrial environment in the development of algorithms. We use a training data set with 95 samples and a test data set with 858 samples. The work uses the contracting process of copper wires to produce formed coil windings to illustrate the method. We analyze a data set with different defect types and compare the quality assessment using the height data acquired with OCT, the feature vectors from the camera images, and the reconstructed height data.

Machine-learning-based (ML) segmentation in the image domain can be utilized for the detection of corrosion on the surface of industrial objects. This research provides a comparison of techniques using convolutional neural networks (CNNs) on the one hand, and random forest (RF) classifiers within RGB and HSV feature spaces on the other hand. CNN-based approaches usually need a large amount of data for training in order for the network to converge and generalize well on new data. Due to the low amount of data provided, we apply a set of methods to increase the generalization ability of the model. These methods can be categorized into data augmentation, selection of larger and smaller models and pretraining strategies like self supervised learning (SSL). The RF classifiers on the other hand are trained per pixel, so that the amount of data is determined by the image size. The object to be tested is a barrel made of metal, from which the image of the coat is used as the training data, and the image of the bottom as test data. We found that a RF classifier in the RGB feature space outperforms the CNNs by seven percentage points regarding the f 1 -score of the corrosion class.