Blink detection and classification can provide a very useful clinical indicator, because of its relation with many neurological and ophthalmological conditions. In this work, we propose a system that automatically detects and classifies blinks as "complete" or "incomplete" in high resolution image sequences zoomed into the participants' face, acquired during clinical examination using near-Infrared illumination. This method utilizes state-of-the-art (DeepLabv3+) deep learning encoder-decoder neural architecture-DLED to segment iris and eyelid in both eyes in the acquired images. The sequence of the segmented frames is post-processed to calculate the distance between the eyelids of each eye (palpebral fissure height) and the corresponding iris diameter. Theses quantities are temporally filtered and their fraction is subject to adaptive thresholding to identify blinks and determine their type, independently for each eye. The proposed system was tested on 15 participants, each with one video of 4 to 10 minutes. Several metrics of blink detection and classification accuracy were calculated against the ground truth, which was generated by three (3) independent experts, whose conflicts were resolved by a senior expert. Results show that the proposed system achieved F1-score 95.3% and 80.9% for the classification of complete and incomplete blinks respectively, collectively for all 15 participants, outperforming all 3 experts. The proposed system was proven robust in handling unexpected participant movements and actions, as well as glare and reflections from the spectacles, or face obstruction by facemasks.

Estimating the homography matrix between images captured under radically different camera poses and zoom factors is a complex challenge. Traditional methods rely on the Random Sample Consensus (RANSAC) algorithm, which requires pairs of homologous points, pre-matched based on local image feature vectors. Sampling consensus is a core step in many Artificial Intelligence (AI) algorithms that enable computer systems to recognize patterns in data. In this paper, we propose H-RANSAC, an algorithm for homography estimation that eliminates the need for feature vectors or explicit point pairing, while it optionally supports point labeling into two classes. H-RANSAC introduces a novel geometric (cheiral) criterion that intelligently rejects implausible point configurations at the beginning of each iteration, while leveraging concave quadrilaterals typically discarded by similar algorithms. A post-hoc criterion at the end of each iteration improves accuracy further. Analytical derivations of the expected maximum iterations are provided, considering success probabilities and outlier rates, enabling adaptive performance tuning. The algorithm is validated on a demanding task: estimating homography between video frames of football matches captured by 12 cameras with highly divergent viewpoints. Results show that H-RANSAC significantly outperforms state-of-the-art classical methods, combined with deep learning-based salient point detection, in terms of average reprojection error and success rates. The relevant implementation is available in https://github.com/gnousias/H-RANSAC.

Blink detection is considered a useful indicator both for clinical conditions and drowsiness state. In this work, we propose and compare deep learning architectures for the task of detecting blinks in video frame sequences. The first step is the training and application of an eye detector that extracts the eye regions from each video frame. The cropped eye regions are organized as three-dimensional (3D) input with the third dimension spanning time of 300 ms. Two different 3D convolutional neural networks are utilized (a simple 3D CNN and 3D ResNet), as well as a 3D autoencoder combined with a classifier coupled to the latent space. Finally, we propose the usage of a frame prediction accumulator combined with morphological processing and watershed segmentation to detect blinks and determine their start and stop frame in previously unseen videos. The proposed framework was trained on ten (9) different participants and tested on five (8) different ones, with a total of 162,400 frames and 1172 blinks for each eye. The start and end frame of each blink in the dataset has been annotate by specialized ophthalmologist. Quantitative comparison with state-of-the-art blink detection methodologies provide favorable results for the proposed neural architectures coupled with the prediction accumulator, with the 3D ResNet being the best as well as the fastest performer.