Criminal investigations oftentimes need the identification of license plates of escape vehicles. The vehicles may be recorded by low-quality cameras in the wild. Their license plates may be unreadable for police officers. Recent efforts aim to use machine learning to forensically decipher license plates from such low-quality images. These methods operate near the information-theoretic limit of recognition and hence show quite high error rates. Unfortunately, it is unclear when such prediction errors occur, which makes it difficult to use these methods in practice. In this work, we propose a Bayesian Neural Network to inherently incorporate a reliability measure into the classifier. We additionally propose to integrate multiple estimations with an entropy weight to further improve the reliability. Our experiments show that this uncertainty metric dramatically reduces the number of false predictions while preserving most of the true predictions.

The demand for high-resolution point clouds has increased throughout the last years. However, capturing high-resolution point clouds is expensive and thus, frequently replaced by upsampling of low-resolution data. Most state-of-the-art methods are either restricted to a rastered grid, incorporate normal vectors, or are trained for a single use case. We propose to use the frequency selectivity principle, where a frequency model is estimated locally that approximates the surface of the point cloud. Then, additional points are inserted into the approximated surface. Our novel frequency-selective geometry upsampling shows superior results in terms of subjective as well as objective quality compared to state-of-the-art methods for scaling factors of 2 and 4. On average, our proposed method shows a 4.4 times smaller point-to-point error than the second best state-of-the-art PU-Net for a scale factor of 4.

Forensic license plate recognition (FLPR) remains an open challenge in legal contexts such as criminal investigations, where unreadable license plates (LPs) need to be deciphered from highly compressed and/or low resolution footage, e.g., from surveillance cameras. In this work, we propose a side-informed Transformer architecture that embeds knowledge on the input compression level to improve recognition under strong compression. We show the effectiveness of Transformers for license plate recognition (LPR) on a low-quality real-world dataset. We also provide a synthetic dataset that includes strongly degraded, illegible LP images and analyze the impact of knowledge embedding on it. The network outperforms existing FLPR methods and standard state-of-the art image recognition models while requiring less parameters. For the severest degraded images, we can improve recognition by up to 8.9 percent points.1

Neural networks are widely used for almost any task of recognizing image content. Even though much effort has been put into investigating efficient network architectures, optimizers, and training strategies, the influence of image interpolation on the performance of neural networks is not well studied. Furthermore, research has shown that neural networks are often sensitive to minor changes in the input image leading to drastic drops of their performance. Therefore, we propose the use of keypoint agnostic frequency selective mesh-to-grid resampling (FSMR) for the processing of input data for neural networks in this paper. This model-based interpolation method already showed that it is capable of outperforming common interpolation methods in terms of PSNR. Using an extensive experimental evaluation we show that depending on the network architecture and classification task the application of FSMR during training aids the learning process. Furthermore, we show that the usage of FSMR in the application phase is beneficial. The classification accuracy can be increased by up to 4.31 percentage points for ResNet50 and the Oxtlower17 dataset.

Neural networks became the standard technique for image classification throughout the last years. They are extracting image features from a large number of images in a training phase. In a following test phase, the network is applied to the problem it was trained for and its performance is measured. In this paper, we focus on image classification. The amount of visual data that is interpreted by neural networks grows with the increasing usage of neural networks. Mostly, the visual data is transmitted from the application side to a central server where the interpretation is conducted. If the transmission is disturbed, losses occur in the transmitted images. These losses have to be reconstructed using postprocessing. In this paper, we incorporate the widely applied bilinear and bicubic interpolation and the high-quality reconstruction Frequency-Selective Reconstruction (FSR) for the reconstruction of corrupted images. However, we propose to use Frequency-Selective Mesh-to-Grid Resampling (FSMR) for the joint reconstruction and resizing of corrupted images. The performance in terms of classification accuracy of EfficientNetB0, DenseNet121, DenseNet201, ResNet50 and ResNet152 is examined. Results show that the reconstruction with FSMR leads to the highest classification accuracy for most networks. Average improvements of up to 6.7 percentage points are possible for DenseNet121.

High frame rates are desired in many fields of application. As in many cases the frame repetition rate of an already captured video has to be increased, frame rate up-conversion (FRUC) is of high interest. We conduct a motion compensated approach. From two neighboring frames, the motion is estimated and the neighboring pixels are shifted along the motion vector into the frame to be reconstructed. For displaying, these irregularly distributed mesh pixels have to be resampled onto regularly spaced grid positions. We use the model-based key point agnostic frequency-selective mesh-to-grid resampling (AFSMR) for this task and show that AFSMR works best for applications that contain irregular meshes with varying densities. AFSMR gains up to 3.2 dB in contrast to the already high performing frequency-selective mesh-to-grid resampling (FSMR). Additionally, AFSMR increases the run time by a factor of 11 relative to FSMR.

With the increased use of virtual and augmented reality applications, the importance of point cloud data rises. High-quality capturing of point clouds is still expensive and thus, the need for point cloud super-resolution or point cloud upsampling techniques emerges. In this paper, we propose an interpolation scheme for color upsampling of three-dimensional color point clouds. As a point cloud represents an object’s surface in three-dimensional space, we first conduct a local transform of the surface into a two-dimensional plane. Secondly, we propose to apply a novel Frequency-Selective Mesh-to-Mesh Resampling (FSMMR) technique for the interpolation of the points in 2D. FSMMR generates a model of weighted superpositions of basis functions on scattered points. This model is then evaluated for the final points in order to increase the resolution of the original point cloud. Evaluation shows that our approach outperforms common interpolation schemes. Visual comparisons of the jaguar point cloud underlines the quality of our upsampling results. The high performance of FSMMR holds for various sampling densities of the input point cloud.

Recently, cross-spectral stereo-camera setups found their way from special applications to mass market, especially in smartphones, automotive systems, or drones. In the following, a novel concept is introduced to bring stereo cameras and cross-spectral disparity estimation together. So far, either monomodal stereo algorithms exist that are not suitable for cross-spectral image registration, or structural template matching is applied that achieves a low quality. To overcome these limitations, a technique is proposed to synthesize arbitrary spectral components from widely available color stereo databases, and to retrain mono-modal deep learning methods. In this contribution, the estimation of spectral bands based on random processes is shown together with noise models, which also allow for a robust registration of narrowband components. The theoretical examination is completed by an extensive evaluation, including a self-manufactured cross-spectral camera setup. In comparison to state-of-the-art techniques, the end-point error is on average reduced by a factor of seven.

It has been shown, that high resolution images can be acquired using a low resolution sensor with non-regular sampling. Therefore, post-processing is necessary. In terms of video data, not only the spatial neighborhood can be used to assist the reconstruction, but also the temporal neighbor-hood. A popular and well performing algorithm for this kind of problem is the three-dimensional frequency selective extrapolation (3D-FSE) for which a motion adapted version is introduced in this paper. This proposed extension solves the problem of changing content within the area considered by the 3D-FSE, which is caused by motion within the sequence. Because of this motion, it may happen that regions are emphasized during the reconstruction that are not present in the original signal within the considered area. By that, false content is introduced into the extrapolated sequence, which affects the resulting image quality negatively. The novel extension, presented in the following, incorporates motion data of the sequence in order to adapt the algorithm accordingly, and compensates changing content, resulting in gains of up to 1.75 dB compared to the existing 3D-FSE.