This paper presents a novel bio-inspired edge-oriented approach to perceptual contour extraction. Our method does not rely on segmentation and can unsupervised learn to identify edge points that are readily grouped, without invoking any connecting mechanism, into object boundaries as perceived by human. This goal is achieved by using a dynamic mask to statistically assess the inter-edge relations and probe the principal direction that acts as an edge-grouping cue. The novelty of this work is that the mask, centered at a target pixel and driven by EM algorithm, can iteratively deform and rotate until it covers pixels that best fit the Bayesian likelihood of the binary class w.r.t a target pixel. By creating an effect of enlarging receptive field, contiguous edges of the same object can be identified while suppressing noise and textures, the resulting contour is in good agreement with gestalt laws of continuity, similarity and proximity. All theoretical derivations and parameters updates are conducted under the framework of EM-based Bayesian inference. Issues of stability and parameter uncertainty are addressed. Both qualitative and quantitative comparison with existing approaches proves the superiority of the proposed method in terms of tracking curved contours, noises/texture resilience, and detection of low-contrast contours.

Conventional image entropy merely involves the overall pixel intensity statistics which cannot respond to intensity patterns over spatial domain. However, spatial distribution of pixel intensity is definitely crucial to any biological or computer vision system, and that is why gestalt grouping rules involve using features of both aspects. Recently, the increasing integration of knowledge from gestalt research into visualization-related techniques has fundamentally altered both fields, offering not only new research questions, but also new ways of solving existing issues. This paper presents a Bayesian edge detector called GestEdge, which is effective in detecting gestalt edges, especially useful for forming object boundaries as perceived by human eyes. GestEdge is characterized by employing a directivity-aware sampling window or mask that iteratively deforms to probe or explore the existence of principal direction of sampling pixels; when convergence is reached, the window covers pixels best representing the directivity in compliance with the similarity and proximity laws in gestalt theory. During the iterative process based on the unsupervised Expectation-Minimization (EM) algorithm, the shape of the sampling window is optimally adjusted. Such a deformable window allows us to exploit the similarity and proximity among the sampled pixels. Comparisons between GestEdge and other edge detectors are shown to justify the effectiveness of GestEdge in extracting the gestalt edges.

Anomaly data are commonly seen in either public or private datasets, as they are easily introduced during the labeling/annotation process. This is particularly true for imagery data. We propose a novel data-cleaning approach (touted as ELDC) characterized by an ensemble of base models that can be trained directly with noisy labels and dynamically provide clean data during iterative training. Without the need to define any distance measures as in many classical ensemble methods, ELDC is particularly suitable for handling imagery data. Each base model uses a random subset of the target dataset that may initially contain up to 40% of label errors. Following each training iteration, anomaly data are discriminated against clean ones by a majority voting scheme, and three different types of anomaly (mislabeled, confusing, and outliers) can be identified using a statistical pattern and removed. Such a cycle of train-vote-remove is iterated until a prespecified condition is reached. ELDC is highly flexible as it is not bound to a specific model and allows different transfer-learning configurations, AlexNet, ResNet50, and GoogleNet were tested as based models and trained with various benchmark datasets, the results show at least 92% accuracy in the presence of 40% of noisy labels, outperforming most of the classical cleaning methods.

Global warming and pollution could lead to the destruction of marine habitats and loss of species. The anomalous behavior of underwater creatures can be used as a biometer for assessing the health status of our ocean. Advances in behavior recognition have been driven by the active application of deep learning methods, yet many of them render superior accuracy at the cost of high computational complexity and slow inference. This paper presents a real-time anomalous behavior recognition approach that incorporates a lightweight deep learning model (Lite3D), object detection, and multitarget tracking. Lite3D is characterized in threefold: (1) image frames contain only regions of interest (ROI) generated by an object detector; (2) no fully connected layers are needed, the prediction head itself is a flatten layer of 1 × $${\mathcal{l}}$$l @ 1× 1, $${\mathcal{l}}$$l = number of categories; (3) all the convolution kernels are 3D, except the first layer degenerated to 2D. Through the tracking, a sequence of ROI-only frames is subjected to 3D convolutions for stacked feature extraction. Compared to other 3D models, Lite3D is 50 times smaller in size and 57 times lighter in terms of trainable parameters and can achieve 99% of F1-score. Lite3D is ideal for mounting on ROV or AUV to perform real-time edge computing.

This study addresses the issue of leveraging federated learning to improve data privacy and performance in IVF embryo selection. The EM (Expectation-Maximization) algorithm is incorporated into deep learning models to form a federated learning framework for quality evaluation of blastomere cleavage using two-dimensional images. The framework comprises a server site and several client sites characterized in that each is locally trained with an EM algorithm. Upon the completion of the local EM training, a separate 5-mode mixture distribution is generated for each client, the clients' distribution statics are then uploaded to the server site and aggregated therein to produce a global (sharing) 5-mode distribution. During the inference phase, each client uses image classifiers and an instance segmentor, assisted by the global 5-mode distribution acting as a calibrator to (1) identify the absolute cleavage timing of blastomere, i.e., tPNa, tPNf, t2, t3, t4, t5, t6, t7, and t8, (2) track the cleavage process of blastomeres to detect the irregular cleavage patterns, and (3) assess the symmetry degree of blastomeres. Experimental results show that the proposed method outperforms commercial Time-Lapse Incubators in reducing the average error of timing prediction by twofold. The proposed facilitate frameworks the adaptability and scalability of classifiers and segmentor to data variability associated with patients in different locations or countries.