Currently there is no publicly available adequate dataset that could be used for training Generative Adversarial Networks (GANs) on car images. All available car datasets differ in noise, pose, and zoom levels. Thus, the objective of this work was to create an improved car image dataset that would be better suited for GAN training. To improve the performance of the GAN, we coupled the LSUN and Stanford car datasets. A new merged dataset was then pruned in order to adjust zoom levels and reduce the noise of images. This process resulted in fewer images that could be used for training, with increased quality though. This pruned dataset was evaluated by training the StyleGAN with original settings. Pruning the combined LSUN and Stanford datasets resulted in 2,067,710 images of cars with less noise and more adjusted zoom levels. The training of the StyleGAN on the LSUN-Stanford car dataset proved to be superior to the training with just the LSUN dataset by 3.7% using the Fréchet Inception Distance (FID) as a metric. Results pointed out that the proposed LSUN-Stanford car dataset is more consistent and better suited for training GAN neural networks than other currently available large car datasets.

A few promising solutions for thermal imaging Unexploded Ordnance (UXO) detection were proposed after the start of the military conflict in Ukraine in 2014. At the same time, most of the landmine clearance protocols and practices are based on old, 20th-century technologies. More than 60 countries worldwide are still affected by explosive remnants of war, and new areas are contaminated almost every day. To date, no automated solutions exist for surface UXO detection by using thermal imaging. One of the reasons is also that there are no publicly available data. This research bridges both gaps by introducing an automated UXO detection method, and by publishing thermal imaging data. During a project in Bosnia and Herzegovina in 2019, an organisation, Norwegian People’s Aid, collected data about unexploded ordnances and made them available for this research. Thermal images with a size of 720 × 480 pixels were collected by using an Unmanned Aerial Vehicle at a height of 3 m, thus achieving a very small Ground Sampling Distance (GSD). One of the goals of our research was also to verify if the explosive war remnants’ detection accuracy could be improved further by using Convolutional Neural Networks (CNN). We have experimented with various existing modern CNN architectures for object identification, whereat the YOLOv5 model was selected as the most promising for retraining. An eleven-class object detection problem was solved primarily in this study. Our data were annotated semi-manually. Five versions of the YOLOv5 model, fine-tuned with a grid-search, were trained end-to-end on randomly selected 640 training and 80 validation images from our dataset. The trained models were verified on the remaining 88 images from our dataset. Objects from each of the eleven classes were identified with more than 90% probability, whereat the Mean Average Precision (mAP) at a 0.5 threshold was 99.5%, and the mAP at thresholds from 0.5 to 0.95 was 87.0% up to 90.5%, depending on the model’s complexity. Our results are comparable to the state-of-the-art, whereat these object detection methods have been tested on other similar small datasets with thermal images. Our study is one of the few in the field of Automated UXO detection by using thermal images, and the first that solves the problem of identifying more than one class of objects. On the other hand, publicly available thermal images with a relatively small GSD will enable and stimulate the development of new detection algorithms, where our method and results can serve as a baseline. Only really accurate automatic UXO detection solutions will help to solve one of the least explored worldwide life-threatening problems.

We evaluated different muscle excitation estimation techniques, and their sensitivity to Motor Unit (MU) distribution in muscle tissue. For this purpose, the Convolution Kernel Compensation (CKC) method was used to identify the MU spike trains from High-Density ElectroMyoGrams (HDEMG). Afterwards, Cumulative MU Spike Train (CST) was calculated by summing up the identified MU spike trains. Muscle excitation estimation from CST was compared to the recently introduced Cumulative Motor Unit Activity Index (CAI) and classically used Root-Mean-Square (RMS) amplitude envelop of EMG. To emphasize their dependence on the MU distribution further, all three muscle excitation estimates were used to calculate the agonist-antagonist co-activation index. We showed on synthetic HDEMG that RMS envelopes are the most sensitive to MU distribution (10 % dispersion around the real value), followed by the CST (7 % dispersion) and CAI (5 % dispersion). In experimental HDEMG from wrist extensors and flexors of post-stroke subjects, RMS envelopes yielded significantly smaller excitations of antagonistic muscles than CST and CAI. As a result, RMS-based co-activation estimates differed significantly from the ones produced by CST and CAI, illuminating the problem of large diversity of muscle excitation estimates when multiple muscles are studied in pathological conditions. Similar results were also observed in experimental HDEMG of six intact young males.

Accurate automated localization of cephalometric landmarks in skull X-ray images is the basis for planning orthodontic treatments, predicting skull growth, or diagnosing face discrepancies. Such diagnoses require as many landmarks as possible to be detected on cephalograms. Today’s best methods are adapted to detect just 19 landmarks accurately in images varying not too much. This paper describes the development of the SCN-EXT convolutional neural network (CNN), which is designed to localize 72 landmarks in strongly varying images. The proposed method is based on the SpatialConfiguration-Net network, which is upgraded by adding replications of the simpler local appearance and spatial configuration components. The CNN capacity can be increased without increasing the number of free parameters simultaneously by such modification of an architecture. The successfulness of our approach was confirmed experimentally on two datasets. The SCN-EXT method was, with respect to its effectiveness, around 4% behind the state-of-the-art on the small ISBI database with 250 testing images and 19 cephalometric landmarks. On the other hand, our method surpassed the state-of-the-art on the demanding AUDAX database with 4695 highly variable testing images and 72 landmarks statistically significantly by around 3%. Increasing the CNN capacity as proposed is especially important for a small learning set and limited computer resources. Our algorithm is already utilized in orthodontic clinical practice.

Automated detection of ovarian follicles in ultrasound images is much appreciated when its effectiveness is comparable with the experts’ annotations. Today’s best methods estimate follicles notably worse than the experts. This paper describes the development of two-stage deeply-supervised 3D Convolutional Neural Networks (CNN) based on the established U-Net. Either the entire U-Net or specific parts of the U-Net decoder were replicated in order to integrate the prior knowledge into the detection. Methods were trained end-to-end by follicle detection, while transfer learning was employed for ovary detection. The USOVA3D database of annotated ultrasound volumes, with its verification protocol, was used to verify the effectiveness. In follicle detection, the proposed methods estimate follicles up to 2.9% more accurately than the compared methods. With our two-stage CNNs trained by transfer learning, the effectiveness of ovary detection surpasses the up-to-date automated detection methods by about 7.6%. The obtained results demonstrated that our methods estimate follicles only slightly worse than the experts, while the ovaries are detected almost as accurately as by the experts. Statistical analysis of 50 repetitions of CNN model training proved that the training is stable, and that the effectiveness improvements are not only due to random initialisation. Our deeply-supervised 3D CNNs can be adapted easily to other problem domains.