Deep Ones Research Articles

Marine compressed air source array acoustic field characterization from at-sea measurements: Long-range propagationa).

In 2007, the Littoral Acoustic Demonstration Center conducted a comprehensive experiment in the northern Gulf of Mexico to measure the three-dimensional acoustic field of a standard marine compressed-air source array used in seismic exploration. This study aims to enhance understanding of long-range acoustic propagation of the array signals, with focus on variations in received sound pressure levels and sound exposure levels (SELs) at various ranges from the source. These variations are influenced by factors, such as receiver depth, array orientation, and propagation conditions. The long-range measurements show that received peak pressure levels and SELs exhibit non-monotonic (oscillatory) behavior with range leading up to 10 dB increase in received levels at longer ranges. At ranges beyond 20 km, acoustic levels at the shallowest hydrophone consistently surpassed those at deeper ones by 3-10 dB, suggesting the impact of surface duct propagation effects. The results demonstrate that range-independent bathymetry leads to approximately 4 dB higher received acoustic levels than range-dependent propagation conditions. The measured long-range propagation acoustic metrics from controlled experiment provide a unique and critical dataset for validating both source and propagation model accuracy in predicting received sound pressure and SEL in the far-field of the source array.

Probabilistic Attention Based on Gaussian Processes for Deep Multiple Instance Learning.

Multiple instance learning (MIL) is a weakly supervised learning paradigm that is becoming increasingly popular because it requires less labeling effort than fully supervised methods. This is especially interesting for areas where the creation of large annotated datasets remains challenging, as in medicine. Although recent deep learning MIL approaches have obtained state-of-the-art results, they are fully deterministic and do not provide uncertainty estimations for the predictions. In this work, we introduce the attention Gaussian process (AGP) model, a novel probabilistic attention mechanism based on Gaussian processes (GPs) for deep MIL. AGP provides accurate bag-level predictions as well as instance-level explainability and can be trained end-to-end. Moreover, its probabilistic nature guarantees robustness to overfit on small datasets and uncertainty estimations for the predictions. The latter is especially important in medical applications, where decisions have a direct impact on the patient's health. The proposed model is validated experimentally as follows. First, its behavior is illustrated in two synthetic MIL experiments based on the well-known MNIST and CIFAR-10 datasets, respectively. Then, it is evaluated in three different real-world cancer detection experiments. AGP outperforms state-of-the-art MIL approaches, including deterministic deep learning ones. It shows a strong performance even on a small dataset with less than 100 labels and generalizes better than competing methods on an external test set. Moreover, we experimentally show that predictive uncertainty correlates with the risk of wrong predictions, and therefore it is a good indicator of reliability in practice. Our code is publicly available.

Clinical Anatomy of the Lower Extremity Veins—Topography, Embryology, Anatomical Variability, and Undergraduate Educational Challenges

Veins of the lower extremity can be categorized into three hierarchically ordered groups: the epifascial, the interfascial, and the deep ones. In the past, the interfascial veins, e.g., the great saphenous vein, were categorized as superficial veins. But nowadays, experts recommend regarding these veins as a separate group because of their unique topography and clinical relevance. In order to better understand the venous anatomy of the lower limbs, which is highly variable, one should also comprehend their embryological development. Venous embryogenesis in the lower limb consists of three stages. During the first stage the primitive fibular vein is the main vein of the extremity. During the second stage it is replaced by the axial vein and finally by the femoral vein. In some adult individuals this embryonic or fetal venous anatomy is still present. Unfortunately, current anatomical textbooks and atlases, as well as traditional cadaver dissections, are not very useful regarding these issues. Therefore, undergraduate teaching of anatomy can be challenging. New educational tools, such as ultrasonography, seem indispensable to teach the anatomy of these veins properly.

Carcinogenic and Non-Carcinogenic Health Risk Evaluation of Heavy Metals in Water Sources of the Nubian Sandstone Aquifer in the El-Farafra Oasis (Egypt)

Expanding anthropogenic activities, globally and in Egypt, have increased concentrations of heavy metals in surface and ground waters. Contamination of drinking water may threaten public health. In the present study, the concentrations of 10 heavy metals were analyzed from natural springs (6) and drilled wells (10) in the Nubian Sandstone Aquifer of the El-Farafra Oasis and the White Desert National Park, Egypt. The average concentrations of heavy metals were in most cases below critical values of the WHO drinking water standard, except for Fe and Mn (average values were 495 and 107 µg·L−1, respectively). There is a surface circulation that develops within limestone (Post-Nubian Aquifer System—PNAS) and feeds the springs, while the water present in the wells (at least for the deeper ones) comes from the ferruginous sandstone (Nubian Sandstone Aquifer System—NSAS). This double circulation could account for the differences in the EC and TDS values (typical of a circulation in limestone-type aquifers for springs) and the Fe and Mn enrichment coming from the ferruginous sandstone of the NSAS. The average chronic daily intake (CDI) values for heavy metals in the study area are listed in decreasing order in the following: Fe > Mn > Zn > Co > Ni > Cr > As > Pb > Co > Cd. The total hazard quotient (HQtotal = HQoral + HQdermal) and Hazard Index (HI) values calculated for different heavy metals were well below the acceptable limit, indicating no significant non-carcinogenic health risks to the residents of both areas via oral and dermal absorption of drinking water. Furthermore, the results obtained for the total risk to human health showed that oral ingestion is the major pathway. Carcinogenic risk analysis indicated that the Incremental Lifetime Cancer Risk (ILCR) values for Pb, Cd, Ni, and Cr were well below the acceptable limits.

Higher-Order Convolutional Neural Networks for Essential Climate Variables Forecasting

Earth observation imaging technologies, particularly multispectral sensors, produce extensive high-dimensional data over time, thus offering a wealth of information on global dynamics. These data encapsulate crucial information in essential climate variables, such as varying levels of soil moisture and temperature. However, current cutting-edge machine learning models, including deep learning ones, often overlook the treasure trove of multidimensional data, thus analyzing each variable in isolation and losing critical interconnected information. In our study, we enhance conventional convolutional neural network models, specifically those based on the embedded temporal convolutional network framework, thus transforming them into models that inherently understand and interpret multidimensional correlations and dependencies. This transformation involves recasting the existing problem as a generalized case of N-dimensional observation analysis, which is followed by deriving essential forward and backward pass equations through tensor decompositions and compounded convolutions. Consequently, we adapt integral components of established embedded temporal convolutional network models, like encoder and decoder networks, thus enabling them to process 4D spatial time series data that encompass all essential climate variables concurrently. Through the rigorous exploration of diverse model architectures and an extensive evaluation of their forecasting prowess against top-tier methods, we utilize two new, long-term essential climate variables datasets with monthly intervals extending over four decades. Our empirical scrutiny, particularly focusing on soil temperature data, unveils that the innovative high-dimensional embedded temporal convolutional network model-centric approaches markedly excel in forecasting, thus surpassing their low-dimensional counterparts, even under the most challenging conditions characterized by a notable paucity of training data.

Coral Shoals Detection from Optical Satellite Imagery Using Deep Belief Network Algorithm: A Case Study for the Xisha Islands, South China Sea

Coral islands and reefs are formed by the cementation of the remains of shallow water reef-building coral polyps and other reef dwelling organisms in tropical oceans. They can be divided into coral islands, coral sandbanks, coral reefs, and coral shoals, of which, Coral shoals are located below the depth datum and are not exposed even at low tide, and sometimes are distributed at water depths exceeding 30 m. Satellite images with wide spatial–temporal coverage have played a crucial role in coral island and reef monitoring, and remote sensing data with multiple platforms, sensors, and spatial and spectral resolutions are employed. However, the accurate detection of coral shoals remains challenging mainly due to the depth effect, that is, coral shoals, especially deeper ones, have very similar spectral characteristics to the sea in optical images. Here, an optical remote sensing detection method is proposed to rapidly and accurately detect the coral shoals using a deep belief network (DBN) from optical satellite imagery. The median filter is used to filter the DBN classification results, and the appropriate filtering window is selected according to the spatial resolution of the optical images. The proposed method demonstrated outstanding performance by validating and comparing the detection results of the Yinli Shoal. Moreover, the expected results are obtained by applying this method to other coral shoals in the Xisha Islands, including the Binmei Shoal, Beibianlang, Zhanhan Shoal, Shanhudong Shoal, and Yongnan Shoal. This detection method is expected to provide the coral shoals’ information rapidly once optical satellite images are available and cloud cover and tropical cyclones are satisfactory. The further integration of the detection results of coral shoals with water depth and other information can effectively ensure the safe navigation of ships.

Synthetic Data for Video Surveillance Applications of Computer Vision: A Review

In recent years, there has been a growing interest in synthetic data for several computer vision applications, such as automotive, detection and tracking, surveillance, medical image analysis and robotics. Early use of synthetic data was aimed at performing controlled experiments under the analysis by synthesis approach. Currently, synthetic data are mainly used for training computer vision models, especially deep learning ones, to address well-known issues of real data, such as manual annotation effort, data imbalance and bias, and privacy-related restrictions. In this work, we survey the use of synthetic training data focusing on applications related to video surveillance, whose relevance has rapidly increased in the past few years due to their connection to security: crowd counting, object and pedestrian detection and tracking, behaviour analysis, person re-identification and face recognition. Synthetic training data are even more interesting in this kind of application, to address further, specific issues arising, e.g., from typically unconstrained image or video acquisition conditions and cross-scene application scenarios. We categorise and discuss the existing methods for creating synthetic data, analyse the synthetic data sets proposed in the literature for each of the considered applications, and provide an overview of their effectiveness as training data. We finally discuss whether and to what extent the existing synthetic data sets mitigate the issues of real data, highlight existing open issues, and suggest future research directions in this field.

Self-Regulated Learning Strategies Use in an Asian High School EFL Context

Objective: This study investigated the self-regulated learning (SRL) strategies employed by high school EFL learners in the Asian high school context and the associations between SRL strategies and their academic achievements. Methods: A total of 6 students in Taiwan region with different academic performances participated in this study to elicit their self-learning experiences and self-regulated strategy use. The data were analyzed using the grounded theory approach by undergoing three stages: open, axial, and selective coding. The comparisons among the data and categories were repeated until no new properties or relationships emerged during the analysis, ensuring theoretical saturation. In the final selective coding stage, two themes emerged from the data sets: (1) degree of teacher directing (2) higher and lower achievers’ SRL strategies use. Results: With different degrees of teacher directing, the pattern of using SRL strategies between higher and lower achievers was revealed. Higher achievers were found to be more autonomous in adopting SRL strategies. Although lower achievers also adopt some SRL strategies, they tend to use surface cognitive strategies, including repetitive rehearsal and rote memorization, instead of deep ones, such as elaboration and organization.Furthermore, in terms of self-consequence, the result also showed that only higher achievers would reward themselves after accomplishing a certain goal and resume their learning. As for low achievers, they tend to give up learning and lose motivation after rewarding themselves. Conclusion: The findings showed that EFL participants’ SRL strategy use varied in response to the self-learning context. Pedagogical implications for future research and how EFL teachers can promote students’ SRL strategy use and English teacher training are discussed.

Contribution to pulmonary diseases diagnostic from X-ray images using innovative deep learning models

Pulmonary disease identification and characterization are among the most intriguing research topics of recent years since they require an accurate and prompt diagnosis. Although pulmonary radiography has helped in lung disease diagnosis, the interpretation of the radiographic image has always been a major concern for doctors and radiologists to reduce diagnosis errors. Due to their success in image classification and segmentation tasks, cutting-edge artificial intelligence techniques like machine learning (ML) and deep learning (DL) are widely encouraged to be applied in the field of diagnosing lung disorders and identifying them using medical images, particularly radiographic ones. For this end, the researchers are concurring to build systems based on these techniques in particular deep learning ones. In this paper, we proposed three deep-learning models that were trained to identify the presence of certain lung diseases using thoracic radiography. The first model, named “CovCXR-Net”, identifies the COVID-19 disease (two cases: COVID-19 or normal). The second model, named “MDCXR3-Net”, identifies the COVID-19 and pneumonia diseases (three cases: COVID-19, pneumonia, or normal), and the last model, named “MDCXR4-Net”, is destined to identify the COVID-19, pneumonia and the pulmonary opacity diseases (4 cases: COVID-19, pneumonia, pulmonary opacity or normal). These models have proven their superiority in comparison with the state-of-the-art models and reached an accuracy of 99,09 %, 97.74 %, and 90,37 % respectively with three benchmarks.

An improved deep hashing model for image retrieval with binary code similarities

The exponential growth of data raises an unprecedented challenge in data analysis: how to retrieve interesting information from such large-scale data. Hash learning is a promising solution to address this challenge, because it may bring many potential advantages, such as extremely high efficiency and low storage cost, after projecting high-dimensional data to compact binary codes. However, traditional hash learning algorithms often suffer from the problem of semantic inconsistency, where images with similar semantic features may have different binary codes. In this paper, we propose a novel end-to-end deep hashing method based on the similarities of binary codes, dubbed CSDH (Code Similarity-based Deep Hashing), for image retrieval. Specifically, it extracts deep features from images to capture semantic information using a pre-trained deep convolutional neural network. Additionally, a hidden and fully connected layer is attached at the end of the deep network to derive hash bits by virtue of an activation function. To preserve the semantic consistency of images, a loss function has been introduced. It takes the label similarities, as well as the Hamming embedding distances, into consideration. By doing so, CSDH can learn more compact and powerful hash codes, which not only can preserve semantic similarity but also have small Hamming distances between similar images. To verify the effectiveness of CSDH, we evaluate CSDH on two public benchmark image collections, i.e., CIFAR-10 and NUS-WIDE, with five classic shallow hashing models and six popular deep hashing ones. The experimental results show that CSDH can achieve competitive performance to the popular deep hashing algorithms.

A review of machine learning-based methods for predicting drug-target interactions.

The prediction of drug-target interactions (DTI) is a crucial preliminary stage in drug discovery and development, given the substantial risk of failure and the prolonged validation period associated with in vitro and in vivo experiments. In the contemporary landscape, various machine learning-based methods have emerged as indispensable tools for DTI prediction. This paper begins by placing emphasis on the data representation employed by these methods, delineating five representations for drugs and four for proteins. The methods are then categorized into traditional machine learning-based approaches and deep learning-based ones, with a discussion of representative approaches in each category and the introduction of a novel taxonomy for deep neural network models in DTI prediction. Additionally, we present a synthesis of commonly used datasets and evaluation metrics to facilitate practical implementation. In conclusion, we address current challenges and outline potential future directions in this research field.

Usability of eye trackers as tools for designers of anastylosis

Eye trackers are more and more often employed by scientists willing to learn more about how cultural heritage is perceived. However, designers very seldom make use of this technology, primarily perhaps due to its expensive and time-consuming nature. By choosing not to employ eye trackers they limit themselves to their own ideas and choices, at the same time forfeiting all the potential advantages of social consultations with non-professionals. One can easily imagine that their work and its effects would only benefit should a way be found to make eye-tracking tests more logistically and financially feasible or if another measure was established that would make it possible to learn the visual reactions of regular people. This paper focuses on perception of anastylosis. There are numerous types of damaged structures that get reassembled and multiple ways of doing that. While most doctrinal documents assert that the original object and the added elements should be easily distinguishable, the two solutions, most often chosen by designers is to make the cavity fillings differ in either texture or luminance from the original material. It is obvious that this results in a large range of solutions to choose from – from shallow indentations in the stone surface to deep ones, from stone just a shade brighter than the original to a distinct cream-white one. Since it is impossible to test all the possible variations using eye trackers, the authors of this paper tried a different approach. Eleven different virtual images were prepared of the same reassembled ionic column. They varied in the level of modifications applied to either the texture or the luminance of the cavity fillings. The entire set of eleven stimuli was then shown to a group of professionals in the field of architecture and/or conservation of monuments. They were asked to choose the best stimuli, that is those that facilitate the differentiation of the old and new parts of the structure by non-professionals. Thus, reduced set of three stimuli was then used in an eye-tracking tests involving over 100 participants. The eye-tracking data, combined with the results of previous research on the perception of anastylosis allowed the authors to formulate several conclusions. It turned out that the images selected by the professionals were far from those that would potentially stimulate and help non-professionals the most. However, the obtained information allowed formulating a few basic guidelines as to the use of differences in texture and luminance in anastylosis. It also made it possible to find a financially and logistically feasible alternative to the demanding eye-tracking tests when it comes to introducing social consultations into the designing process.

The influence of driftwood debris in accelerating the regional scour near the bridge pier

Flood disasters caused by driftwood debris are increasing and becoming more severe due to rising rainfall, shifting forest conditions, and increased trees along rivers. Driftwood buildup in rivers leads to sedimentation, erosion, and rising water levels, necessitating precise prediction of its behavior to mitigate its harmful effects. This study utilizes a three-dimensional numerical model to investigate the influence of piled-up driftwood debris on the acceleration of local scour development in front of the bridge pier. Depending on their density, driftwood blockages can be deposition types, such as floating, suspended, and submerged materials. The floating driftwood debris types yielded broader and deeper scour holes, while the suspended driftwood debris produced narrower but deeper ones. The submerged driftwood debris obstructed the pier perimeter; consequently, the plunging and spiral horseshoe patterns deteriorated as the flow approached the pier. The horseshoe vortex’s intensity and diving flow are likely lower, producing shallower local erosion around the bridge pier.

Investigation on the influence of mixed borehole depths in vertical ground heat exchanger systems

Most studies about vertical borehole heat exchangers (BHEs) tended to focus on optimizing total borehole length with all boreholes having the same depth. This is mainly because most of the vertical BHEs design tools do not have the feature of designing mixed borehole depths at the same bore field. Because thermal saturation occurs over time at the center of the field due to thermal interaction between the confined boreholes in the interior region, reducing depth of these less efficient boreholes is a desired goal. This paper studies the effects of mixed borehole depths where shorter boreholes are designed to be in the interior region, and deeper ones in the perimeter. The python package pygfunction has been adapted in this paper to generate the thermal response factor of a bore field with mixed borehole depths and to simulate its performance in difference design scenarios. The results reveal that total BHE-length could be decreased by up to 10 % in some cases of mixed borehole depths compared with the case of same-depth boreholes. However, since the total BHE-length could increase in few cases, careful examination is crucial when applying mixed borehole depths.

Deep quantum graph dreaming: deciphering neural network insights into quantum experiments

Despite their promise to facilitate new scientific discoveries, the opaqueness of neural networks presents a challenge in interpreting the logic behind their findings. Here, we use a eXplainable-AI technique called inception or deep dreaming, which has been invented in machine learning for computer vision. We use this technique to explore what neural networks learn about quantum optics experiments. Our story begins by training deep neural networks on the properties of quantum systems. Once trained, we ‘invert’ the neural network—effectively asking how it imagines a quantum system with a specific property, and how it would continuously modify the quantum system to change a property. We find that the network can shift the initial distribution of properties of the quantum system, and we can conceptualize the learned strategies of the neural network. Interestingly, we find that, in the first layers, the neural network identifies simple properties, while in the deeper ones, it can identify complex quantum structures and even quantum entanglement. This is in reminiscence of long-understood properties known in computer vision, which we now identify in a complex natural science task. Our approach could be useful in a more interpretable way to develop new advanced AI-based scientific discovery techniques in quantum physics.

Subsurface characterization by active and passive source geophysical methods after the 06 February 2023 earthquakes in Turkey

Two large earthquakes (Mw = 7.7 and Mw = 7.6) that occurred in Turkey on February 6, 2023, affected a very extent region and caused a lot of loss of life and property. This paper presents preliminary results from geophysical measurements (Seismic Refraction Tomography-SRT, Multi-Channel Surface Wave Analysis-MASW and Microtremor-MT) on eight profiles in four provinces (Kahramanmaraş, Hatay, Malatya, Gaziantep) to understand the relationship between subsurface properties and the destruction that occurs immediately after earthquakes. By analyzing the geophysical data, the dynamic-elastic properties of ground and the soil classification according to Vs30 were determined. It is generally understood that the near-surface (< ~ 10–15 m) units in the measurement areas are very loose, and the deeper ones (≥ ~ 15–20 m) have a very porous/fractured structure. Soil classes were defined as ZD (Malatya-1, Hatay-1 and Kahramanmaraş-1) and ZC (Malatya-2, Hatay-2, Gaziantep-1,2 and Kahramanmaraş-2). In addition, by evaluating the information of strong ground motion station closest to the measurement profiles, it is observed that the PGA values versus epicenter distances are higher at stations in the zone parallel to the direction of both faults than those in the perpendicular zones. This leads directivity effect in the propagation of earthquake waves. The results indicate that one of the basic reasons for the damages is that the earthquake-ground-structure relationship has not been fully and accurately reflected in building designs. Therefore, future researches involving more geophysical data and PGA values will provide more information about the structural, physical and geotechnical properties of subsurface and definitive results.

Spectral Investigation of the Relationship between Seismicity and Water Level in the Enguri High Dam Area (Georgia)

A spectral analysis of the time dynamics of seismicity occurring in the Enguri area of Georgia from 1978 to 2021 is performed by means of Schuster’s spectrum analysis, periodogram analysis, and empirical mode decomposition. The results of our analysis suggest that earthquakes around the reservoir (within a 50 km radius from the center of the dam) may be due to changes in water level, featured by the yearly cycle of loading and unloading operations of the reservoir. It is observed that the impacts of water fluctuations are more pronounced in shallower strata (down to 10 km) than deeper ones (down to 20 km); this could indicate that earthquakes occurring at deeper levels may primarily result from tectonic forces, whereas those at shallower depths may be predominantly triggered by reservoir-induced factors.

DILS: depth incremental learning strategy.

There exist various methods for transferring knowledge between neural networks, such as parameter transfer, feature sharing, and knowledge distillation. However, these methods are typically applied when transferring knowledge between networks of equal size or from larger networks to smaller ones. Currently, there is a lack of methods for transferring knowledge from shallower networks to deeper ones, which is crucial in real-world scenarios such as system upgrades where network size increases for better performance. End-to-end training is the commonly used method for network training. However, in this training strategy, the deeper network cannot inherit the knowledge from the existing shallower network. As a result, not only is the flexibility of the network limited but there is also a significant waste of computing power and time. Therefore, it is imperative to develop new methods that enable the transfer of knowledge from shallower to deeper networks. To address the aforementioned issue, we propose an depth incremental learning strategy (DILS). It starts from a shallower net and deepens the net gradually by inserting new layers each time until reaching requested performance. We also derive an analytical method and a network approximation method for training new added parameters to guarantee the new deeper net can inherit the knowledge learned by the old shallower net. It enables knowledge transfer from smaller to larger networks and provides good initialization of layers in the larger network to stabilize the performance of large models and accelerate their training process. Its reasonability can be guaranteed by information projection theory and is verified by a series of synthetic and real-data experiments.

Decouple Graph Neural Networks: Train Multiple Simple GNNs Simultaneously Instead of One.

Graph neural networks (GNN) suffer from severe inefficiency due to the exponential growth of node dependency with the increase of layers. It extremely limits the application of stochastic optimization algorithms so that the training of GNN is usually time-consuming. To address this problem, we propose to decouple a multi-layer GNN as multiple simple modules for more efficient training, which is comprised of classical forward training (FT) and designed backward training (BT). Under the proposed framework, each module can be trained efficiently in FT by stochastic algorithms without distortion of graph information owing to its simplicity. To avoid the only unidirectional information delivery of FT and sufficiently train shallow modules with the deeper ones, we develop a backward training mechanism that makes the former modules perceive the latter modules, inspired by the classical backward propagation algorithm. The backward training introduces the reversed information delivery into the decoupled modules as well as the forward information delivery. To investigate how the decoupling and greedy training affect the representational capacity, we theoretically prove that the error produced by linear modules will not accumulate on unsupervised tasks in most cases. The theoretical and experimental results show that the proposed framework is highly efficient with reasonable performance, which may deserve more investigation.

A DEEP LEARNING-BASED APPROACH FOR DETECTING AND CLASSIFYING INAPPROPRIATE CONTENT IN YOUTUBE VIDEOS

There are now billions of videos on YouTube, with most of the viewers being young people. The growth in the quantity of videos has been exponential. To add insult to injury, some unscrupulous users post unsettling animated cartoon films and other material that parents should not allow their children to see. Therefore, social media sites should have a way to filter videos automatically in real time. In order to identify and categorize objectionable video material, this research proposes a new architecture based on deep learning. To do this, the suggested system employs a convolutional neural network (CNN) model named EfficientNet-B7 that has been pre-trained on ImageNet to extract video descriptors. A bidirectional long short-term memory (BiLSTM) network is trained to classify videos into many categories using these characteristics. In order to distribute attention probabilities throughout the network, an attention mechanism is also included after BiLSTM. We test these models using a dataset that has 111,156 cartoon clips tagged by humans from YouTube videos. The experimental findings showed that EfficientNet-BiLSTM achieved a higher accuracy (D=95.30%) than the EfficientNet-BiLSTM framework that was based on the attention mechanism. To continue, when compared to more conventional machine learning classifiers, deep learning ones outperform the pack. With an overall f1 score of D 0.9267, the combination of EfficientNet and BiLSTM, which includes 128 hidden units, achieved state-of-the-art performance. Furthermore, when stacked on top of CNN, BiLSTM outperforms state-of-the-art methods in detecting and classifying children-unsuitable video content because it better grasps the contextual information of video descriptors in network architecture.