Regular Convolutional Neural Networks Research Articles

Optically processed Kannada script realization with Siamese neural network model

<p><span>Optical character recognition (OCR) is a technology that allows computers to recognize and extract text from images or scanned documents. It is commonly used to convert printed or handwritten text into machine-readable format. This Study presents an OCR system on Kannada Characters based on siamese neural network (SNN). Here the SNN, a Deep neural network which comprises of two identical convolutional neural network (CNN) compare the script and ranks based on the dissimilarity. When lesser dissimilarity score is identified, prediction is done as character match. In this work the authors use 5 classes of Kannada characters which were initially preprocessed using grey scaling and convert it to pgm format. This is directly input into the Deep convolutional network which is learnt from matching and non-matching image between the CNN with contrastive loss function in Siamese architecture. The Proposed OCR system uses very less time and gives more accurate results as compared to the regular CNN. The model can become a powerful tool for identification, particularly in situations where there is a high degree of variation in writing styles or limited training data is available.</span></p>

Ultrasensitive Online NO Sensor Based on a Distributed Parallel Self-Regulating Neural Network and Ultraviolet Differential Optical Absorption Spectroscopy for Exhaled Breath Diagnosis.

The concentration of fractional exhaled nitric oxide (FeNO) is closely related to human respiratory inflammation, and the detection of its concentration plays a key role in aiding diagnosing inflammatory airway diseases. In this paper, we report a gas sensor system based on a distributed parallel self-regulating neural network (DPSRNN) model combined with ultraviolet differential optical absorption spectroscopy for detecting ppb-level FeNO concentrations. The noise signals in the spectrum are eliminated by discrete wavelet transform. The DPSRNN model is then built based on the separated multipeak characteristic absorption structure of the UV absorption spectrum of NO. Furthermore, a distributed parallel network structure is built based on each absorption feature region, which is given self-regulating weights and finally trained by a unified model structure. The final self-regulating weights obtained by the model indicate that each absorption feature region contributes a different weight to the concentration prediction. Compared with the regular convolutional neural network model structure, the proposed model has better performance by considering the effect of separated characteristic absorptions in the spectrum on the concentration and breaking the habit of bringing the spectrum as a whole into the model training in previous related studies. Lab-based results show that the sensor system can stably achieve high-precision detection of NO (2.59-750.66 ppb) with a mean absolute error of 0.17 ppb and a measurement accuracy of 0.84%, which is the best result to date. More interestingly, the proposed sensor system is capable of achieving high-precision online detection of FeNO, as confirmed by the exhaled breath analysis.

Detecting phishing websites through improving convolutional neural networks with Self-Attention mechanism

Emerging technologies have made internet connection a vital activity facilitating access to many services. However, internet connection raises many security concerns, such as illegally acquiring private information, passwords, and identifiers. Phishing websites are the first choice for attackers that try to have users' private space. Social engineering attacks are performed by designing fake websites similar to real ones and inviting the victim to access those websites to collect their sensitive information and then redirect them to the actual site. Due to the importance of detecting phishing websites, building a robust detector that filters them and blocks their activity on the Internet is necessary. In this paper, we proposed a phishing website detector based on improving the convolutional neural network (CNN) with a self-attention mechanism. The proposed detector collects phishing Uniform Resource Locators (URLs) by treating them as strings. CNN models have proved their efficiency when dealing with text strings compared to Long Short-Term Memory (LSTM) which focuses on temporal features. Using CNN allows learning comprehensive features of the URLs and facilitates the detection of phishing ones. The self-attention mechanism was added to enhance the model's focus and detection accuracy. Besides, the training dataset was balanced by generating phishing URLs using a Generative Adversarial Network (GAN). A set of experiments has proved the robustness of the proposed detector by achieving high detection accuracy on the test set. Besides, the proposed detector was tested using unknown URLs and achieved excellent results. The improved CNN's detection precision of 99.7 is higher than the regular CNN model by 2.74%. The reported results show that using the self-attention mechanism has improved the detection accuracy and made the CNN model more efficient for detecting phishing websites.

Comparison of CNN-based methods for yoga pose classification

Yoga is an exercise developed in ancient India. People perform yoga in order to have mental, physical, and spiritual benefits. While yoga helps build strength in the mind and body, incorrect postures might result in serious injuries. Therefore, yoga exercisers need either an expert or a platform to receive feedback on their performance. Since access to experts is not an option for everyone, a system to provide feedback on the yoga poses is required. To this end, commercial products such as smart yoga mats and smart pants are produced; Kinect cameras, sensors, and wearable devices are used. However, these solutions are either uncomfortable to wear or not affordable for everyone. Nonetheless, a system that employs computer vision techniques is a requirement. In this paper, we propose a deep-learning model for yoga pose classification, which is the first step of a quality assessment and personalized feedback system. We introduce a wavelet-based model that first takes wavelet transform of input images. The acquired subbands, i.e., approximation, horizontal, vertical, and diagonal coefficients of the wavelet transform are then fed into separate convolutional neural networks (CNN). The obtained probability results for each group are fused to predict the final yoga class. A publicly available dataset with 5 yoga poses is used. Since the number of images in the dataset is not enough for a deep learning model, we also perform data augmentation to increase the number of images. We compare our results to a CNN model and the three models that employ the subbands separately. Results obtained using the proposed model outperforms the accuracy output achieved with the compared models. While the regular CNN model has 61% and 50% accuracy for the training and test data, the proposed model achieves 91% and 80%, respectively.

TinyAD: Memory-Efficient Anomaly Detection for Time-Series Data in Industrial IoT

Monitoring and detecting abnormal events in cyber-physical systems is crucial to industrial production. With the prevalent deployment of the Industrial Internet of Things (IIoT), an enormous amount of time series data is collected to facilitate machine learning models for anomaly detection, and it is of the utmost importance to directly deploy the trained models on the IIoT devices. However, it is most challenging to deploy complex deep learning models such as Convolutional Neural Networks (CNNs) on these memory-constrained IIoT devices embedded with microcontrollers (MCUs). To alleviate the memory constraints of MCUs, we propose a novel framework named Tiny Anomaly Detection (TinyAD) to efficiently facilitate onboard inference of CNNs for real-time anomaly detection. First, we conduct a comprehensive analysis of depthwise separable CNNs and regular CNNs for anomaly detection and find that the depthwise separable convolution operation can reduce the model size by 50-90% compared with the traditional CNNs. Then, to reduce the peak memory consumption of CNNs, we explore two complementary strategies, in-place, and patch-by-patch memory rescheduling, and integrate them into a unified framework. The in-place method decreases the peak memory of the depthwise convolution by sparing a temporary buffer to transfer the activation results, while the patch-by-patch method further reduces the peak memory of layer-wise execution by slicing the input data into corresponding receptive fields and executing in order. Furthermore, by adjusting the dimension of convolution filters, these strategies apply to both univariate time series and multidomain time series features. Extensive experiments on real-world industrial datasets show that our framework can reduce peak memory consumption by 2-5x with negligible computation overhead.

Evaluating the influence of road construction on landslide susceptibility in Saudi Arabia's mountainous terrain: a Bayesian-optimised deep learning approach with attention mechanism and sensitivity analysis.

In the mountainous region of Asir region of Saudi Arabia, road construction activities are closely associated with frequent landslides, posing significant risks to both human life and infrastructural development. This highlights an urgent need for a highly accurate landslide susceptibility map to guide future development and risk mitigation strategies. Therefore, this study aims to (1) develop robust well-optimised deep learning (DL) models for predicting landslide susceptibility and (2) conduct a comprehensive sensitivity analysis to quantify the impact of each parameter influencing landslides. To achieve these aims, three advanced DL models-Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), and Bayesian-optimised CNN with an attention mechanism-were rigorously trained and validated. Model validation included eight matrices, calibration curves, and Receiver Operating Characteristic (ROC) and Precision-Recall curves. Multicollinearity was examined using Variance Inflation Factor (VIF) to ensure variable independence. Additionally, sensitivity analysis was used to interpret the models and explore the influence of parameters on landslide. Results showed that road networks significantly influenced the areas identified as high-risk zones. Specifically, in the 1-km buffer around roadways, CNN_AM identified 10.42% of the area as 'Very High' susceptibility-more than double the 4.04% indicated by DNN. In the extended 2-km buffer zone around roadways, Bayesian CNN_AM continued to flag a larger area as Very High risk (7.46%), in contrast to DNN's 3.07%. In performance metrics, CNN_AM outshined DNN and regular CNN models, achieving near-perfect scores in Area Under the Curve (AUC), precision-recall, and overall accuracy. Sensitivity analysis highlighted 'Soil Texture', 'Geology', 'Distance to Road', and 'Slope' as crucial for landslide prediction. This research offers a robust, high-accuracy model that emphasises the role of road networks in landslide susceptibility, thereby providing valuable insights for planners and policymakers to proactively mitigate landslide risks in vulnerable zones near existing and future road infrastructure.

Beyond CNNs: Exploiting Further Inherent Symmetries in Medical Image Segmentation.

Automatic tumor or lesion segmentation is a crucial step in medical image analysis for computer-aided diagnosis. Although the existing methods based on convolutional neural networks (CNNs) have achieved the state-of-the-art performance, many challenges still remain in medical tumor segmentation. This is because, although the human visual system can detect symmetries in 2-D images effectively, regular CNNs can only exploit translation invariance, overlooking further inherent symmetries existing in medical images, such as rotations and reflections. To solve this problem, we propose a novel group equivariant segmentation framework by encoding those inherent symmetries for learning more precise representations. First, kernel-based equivariant operations are devised on each orientation, which allows it to effectively address the gaps of learning symmetries in existing approaches. Then, to keep segmentation networks globally equivariant, we design distinctive group layers with layer-wise symmetry constraints. Finally, based on our novel framework, extensive experiments conducted on real-world clinical data demonstrate that a group equivariant Res-UNet (called GER-UNet) outperforms its regular CNN-based counterpart and the state-of-the-art segmentation methods in the tasks of hepatic tumor segmentation, COVID-19 lung infection segmentation, and retinal vessel detection. More importantly, the newly built GER-UNet also shows potential in reducing the sample complexity and the redundancy of filters, upgrading current segmentation CNNs, and delineating organs on other medical imaging modalities.

Simulated mental imagery for robotic task planning.

Traditional AI-planning methods for task planning in robotics require a symbolically encoded domain description. While powerful in well-defined scenarios, as well as human-interpretable, setting this up requires a substantial effort. Different from this, most everyday planning tasks are solved by humans intuitively, using mental imagery of the different planning steps. Here, we suggest that the same approach can be used for robots too, in cases which require only limited execution accuracy. In the current study, we propose a novel sub-symbolic method called Simulated Mental Imagery for Planning (SiMIP), which consists of perception, simulated action, success checking, and re-planning performed on 'imagined' images. We show that it is possible to implement mental imagery-based planning in an algorithmically sound way by combining regular convolutional neural networks and generative adversarial networks. With this method, the robot acquires the capability to use the initially existing scene to generate action plans without symbolic domain descriptions, while at the same time, plans remain human-interpretable, different from deep reinforcement learning, which is an alternative sub-symbolic approach. We create a data set from real scenes for a packing problem of having to correctly place different objects into different target slots. This way efficiency and success rate of this algorithm could be quantified.

Robust facial expression recognition with Transformer Block Enhancement Module

Recently, facial expression recognition (FER) methods have achieved significant progress. However, FER is still challenged by factors such as uneven illumination and low-quality expression images. Exploring the potential of facial expression features to achieve robust FER is particularly important. Inspired by the Transformer’s excellent performance in modeling long-range dependencies and in vision tasks, this paper proposes a Transformer Block Enhancement Module (TBEM) for enhancing the feature representation of facial expressions. The proposed module contains a Channel Enhancement (CE) block and a Spatial Enhancement (SE) block. The CE block can adaptively enhance the expression features on the channel dimension by effectively leveraging the channel dependency information, while the SE block enhances the expression features on the spatial dimension by integrating the spatial dependency information. TBEM can output a more robust expression representation by combining CE and SE. The proposed artificial intelligence learning module greatly improves the recognition accuracy of FER engineering tasks. To further illustrate the application of TBEM in real-world FER engineering, three engineering problems are used for verification. Extensive experiments demonstrate that the proposed method improves FER performance by focusing on more accurate decisional features and can be easily embedded into regular convolutional neural network models to help them improve the accuracy on FER tasks by about 2.64%–3.03%. The results show the proposed method achieves accuracies of 90.57% on FERPlus, 89.41% on RAFDB basic and 68.43% on RAFDB compound, respectively. The proposed method also provides a meaningful reference for further research on applying Transformer to other tasks.

An Evolutionary Fake News Detection Based on Tropical Convolutional Neural Networks (TCNNs) Approach

In general, the characteristics of false news are difficult to distinguish from those of legitimate news. Even if it is wrong, people can make money by spreading false information. A long time ago, there were fake news stories, including the one about "Bat-men on the moon" in 1835. A mechanism for fact-checking statements must be put in place, particularly those that garner thousands of views and likes before being refuted and proven false by reputable sources. Many machine learning algorithms have been used to precisely categorize and identify fake news. In this experiment, an ML classifier was employed to distinguish between fake and real news. In this study, we present a Tropical Convolutional Neural Networks (TCNNs) model-based false news identification system. Convolutional neural networks (CNNs), Gradient Boost, long short-term memory (LSTMs), Random Forest, Decision Tree (DT), Ada Boost, and attention mechanisms are just a few of the cutting-edge techniques that are compared in our study. Furthermore, because tropical convolution operators are fundamentally nonlinear operators, we anticipate that TCNNs will be better at nonlinear fitting than traditional CNN. Our analysis leads us to the conclusion that the Tropical Convolutional Neural Networks (TCNNs) model with attention mechanism has the maximum accuracy of 98.93%. The findings demonstrate that TCNN can outperform regular convolutional neural network (CNN) layers in terms of expressive capability.

Oil species identification based on fluorescence excitation-emission matrix and transformer-based deep learning

After oil spills are found at sea, the identification on oil species can help determine the source of leakage and form the plan of post-accident treatment. Since the fluorometric characteristics of petroleum hydrocarbon reflect its molecular structure, the composition of oil spills could potentially be inferred using the fluorescence spectroscopy method. The excitation-emission matrix (EEM) includes additional fluorescence information in the dimension of excitation wavelength, which could be useful to identify oil species. This study proposed an oil species identification model using transformer network. The EEMs of oil pollutants are reconstructed into sequenced patch input that consists of the fluorometric spectra obtained under the different excitation wavelengths. The comparative experiments show that the proposed model can reduce the incorrect predictions and achieve higher identification accuracies than the regular convolutional neural networks that have been used in the previous studies. According to the structure of transformer network, an ablation experiment is also designed to evaluate the contributions of different input patches and seek for the optimal excitation wavelengths for oil species identification. The proposed model is expected to identify oil species, and even other fluorescent materials, based on the fluorometric spectra collected under multiple excitation wavelengths.

Potential of machine learning approaches for predicting mechanical properties of spruce wood in the transverse direction

To predict the mechanical properties of wood in the transverse direction, this study used machine learning to extract the anatomical features of wood from cross-sectional stereograms. Specimens with different orientations of the ray parenchyma cell were prepared, and their modulus of elasticity (MOE) and modulus of rupture (MOR) were measured by a three-point bending test. The orientation of the ray parenchyma cell and wood density (rho ) were used as parameters for the MOE and MOR prediction. Conventional machine learning algorithms and artificial neural network were used, and satisfactory results were obtained in both cases. A regular convolutional neural network (CNN) and a density-informed CNN were used to automatically extract anatomical features from the specimens’ cross-sectional stereograms to predict the mechanical properties. The regular CNN achieved acceptable but relatively low accuracy in both the MOE and MOR prediction. The reason for this may be that rho information could not be satisfactorily extracted from the images, because the images represented a limited region of the specimen. For the density-informed CNN, the average prediction coefficient for both the MOE and MOR drastically increased when rho information was provided. A regression activation map was constructed to understand the representative anatomical features that are strongly related to the prediction of mechanical properties. For the regular CNN, the latewood region was highly activated in both the MOE and MOR prediction. It is believed that the ratio and orientation of latewood were successfully extracted for the prediction of the considered mechanical properties. For the density-informed CNN, the activated region is different. The earlywood region was activated in the MOE prediction, while the transition region between the earlywood and latewood was activated in the MOR prediction. These results may provide new insights into the relationship between the anatomical features and mechanical properties of wood.

Computationally effective estimation of supersonic flow field around airfoils using sparse convolutional neural network

This work proposes an innovative approach for supersonic flow field modeling around airfoils based on sparse convolutional neural networks (SCNNs) and Bézier generative adversarial network (GAN), where (1) the SCNN model is built to end-to-end predict supersonic compressible physical flow fields around airfoils from spatially-sparse geometries and (2) the trained Bézier-GAN is utilized to generate plenty of smooth airfoils as well as the latent codes representing airfoils. The spatially-sparse positions of airfoil geometry are represented using signed distance function (SDF). Particularly, the latent codes are merged with the SDF matrix and the Mach number to form the input of the SCNN model, effectively making the SCNN model possess more robust geometric adaptability to different flow conditions. The most significant contribution compared to the regular convolutional neural network is that SCNN introduces sparse convolutional operations to process spatially-sparse input matrix, specifically, which only focuses on the local area with flow information when performing convolution, eventually saving memory usage and improving the network’s attention on the flow area. Further, the testing results show that the SCNN model can more accurately predict supersonic flow fields with a mean absolute error lower than 5% and save 40% of graphics processing unit memory. These results indicate that the proposed SCNN model can capture the shock wave features of supersonic flow fields and improve learning efficiency and computing efficiency.

Deep neural networks and gradient-weighted class activation mapping to classify and analyze EEG

Deep Neural Networks (DNN) have significantly improved the capabilities for analysis and classification of data, including that of biomedical signals (eg. ElectroencephalogramEEG). Optimal classification of EEG signals from seizure onset zones has been challenging especially given the complexity of signals arising from multiple locations. Also, underlying electrophysiological abnormalities which signify epileptogenic zones have not been clearly defined. Previous studies have demonstrated, automatic feature generation based on deep learning as a useful tool for interictal epileptiform discharge (IEDs) detection. Also signals with transformations have been used in convolutional neural network (CNN) based models in the past for classifying EEG data. We explored the use of deep learning for identification of the seizure onset zones using regular dense neural network and CNN based models. After the training the model using sample data the results were validated with a smaller percentage (10%) of the data. The models were noted to be accurate in predicting the seizure onset zones with significant degree of accuracy (87–99%) with a much shorter duration of signal recorded, compared to previous studies. We further investigated model’s decision-making process with heatmapping (gradient-weighted class activation map: Grad-CAM) approach, combining with signal processing using Hilbert transform. To identify the visible features in the signal, which maximally contribute to the seizure onset zone prediction, correlation of heatmap and analytical signal of the EEG, was calculated. A high correlation between heatmap and analytical signal was noted, suggesting that the model may be utilizing the higher power regions of the data in decision making process. This study demonstrates the potential use of DNN based strategies in identifying the epileptogenic zones in the intracranial EEG. It also shows the heatmapping strategies can help in establishing how specific signal patterns may be contributing to the decision making of DNN.

Classification and regression-based convolutional neural network and long short-term memory configuration for bridge damage identification using long-term monitoring vibration data

Considerable attention has recently been focused on classification and regression-based convolutional neural network (CNN) and long short-term memory (LSTM) due to their excellent performance in capturing complex spatial and temporal information characteristics for structural damage identification. However, few studies have considered structural damage identification as a classification and regression problem. In addition, bridges in practical engineering are vulnerable to various environmental and vehicle loading conditions. Hence, this study proposed a new two-stage CNN–LSTM configuration for bridge damage identification using vibration data considering the influence of temperatures. First, a classification-based CNN–LSTM is designed to perform multiclass damage detection tasks, and then a regression-based CNN–LSTM is developed for damage localization and severity prediction tasks. The performance of the proposed damage identification method was evaluated through a simulation dataset of a concrete highway bridge model and a field experiment dataset of Z24-bridge (Switzerland). In addition, a set of statistical evaluation metrics such as sparse categorical cross-entropy loss, accuracy, confusion matrix, mean squared loss, mean absolute error, mean absolute percentage error, and coefficient of determination were used to compare the damage identification performance of the proposed CNN–LSTM configuration with a regular CNN model and conventional machine learning (ML) algorithms. Prediction results indicate that the proposed CNN–LSTM model outperforms the regular CNN model and conventional ML algorithms for bridge damage identification.

Machine learning based prediction of the corrosion behaviour of biodegradable magnesium alloys with PEO coatings

AbstractThe present work focuses on the application of deep learning methods for predicting the in vitro pitting corrosion behaviour of coated magnesium alloys with different microstructures. The datasets are generated by phase field simulation of pitting corrosion. Performance analysis shows that physics‐informed convolutional neural network (CNN) is superior to regular CNN in predicting corrosion curve. A good agreement between deep learning predictions and the ground truths indicates excellent predictive performance of the current encoder‐decoder architecture regardless of the variation of microstructures. Furthermore, the computational time is reduced by 4 to 6 orders of magnitude compared to the phase field simulation. The computational results demonstrate that machine learning based approach is promising alternative to traditional modelling methods for predicting pitting corrosion of coated magnesium alloys.

Deep convolutional surrogates and freedom in thermal design

A deep learning approach is presented for heat transfer and pressure drop prediction of complex fin geometries generated using composite Bézier curves. Thermal design process includes iterative high fidelity simulation which is complex, computationally expensive, and time-consuming. With the advancement in machine learning algorithms as well as Graphics Processing Units (GPUs), parallel processing architecture of GPUs can be used to accelerate thermo-fluid simulation. In this study, Convolutional Neural Networks (CNNs) are used to predict results of Computational Fluid Dynamics (CFD) directly from topologies saved as images. A design space with a single fin as well as multiple morphable fins are studied. A comparison of Xception network and regular CNN is presented for the case with a single fin design. Results show that high accuracy in prediction is observed for single fin design particularly using Xception network. Xception network provides 98 percent accuracy in heat transfer and pressure drop prediction of the single fin design. Increasing the design freedom to multiple fins increases the error in prediction. This error, however, remains within three percent of the ground truth values which is valuable for design purpose. The presented predictive model can be used for innovative BREP-based fin design optimization in compact and high efficiency heat exchangers.

Classification of COVID-19 from CT chest images using Convolutional Wavelet Neural Network

<p>Analyzing X-rays and computed tomography-scan (CT scan) images using a convolutional neural network (CNN) method is a very interesting subject, especially after coronavirus disease 2019 (COVID-19) pandemic. In this paper, a study is made on 423 patients’ CT scan images from Al-Kadhimiya (Madenat Al Emammain Al Kadhmain) hospital in Baghdad, Iraq, to diagnose if they have COVID or not using CNN. The total data being tested has 15000 CT-scan images chosen in a specific way to give a correct diagnosis. The activation function used in this research is the wavelet function, which differs from CNN activation functions. The convolutional wavelet neural network (CWNN) model proposed in this paper is compared with regular convolutional neural network that uses other activation functions (exponential linear unit (ELU), rectified linear unit (ReLU), Swish, Leaky ReLU, Sigmoid), and the result is that utilizing CWNN gave better results for all performance metrics (accuracy, sensitivity, specificity, precision, and F1-score). The results obtained show that the prediction accuracies of CWNN were 99.97%, 99.9%, 99.97%, and 99.04% when using wavelet filters (rational function with quadratic poles (RASP1), (RASP2), and polynomials windowed (POLYWOG1), superposed logistic function (SLOG1)) as activation function, respectively. Using this algorithm can reduce the time required for the radiologist to detect whether a patient has COVID or not with very high accuracy.</p>

A deformable CNN architecture for predicting clinical acceptability of ECG signal

The degraded quality of the electrocardiogram (ECG) signals is the main source of false alarms in critical care units. Therefore, a preliminary analysis of the ECG signal is required to decide its clinical acceptability. In conventional techniques, different handcrafted features are extracted from the ECG signal based on signal quality indices (SQIs) to predict clinical acceptability. A one-dimensional deformable convolutional neural network (1D-DCNN) is proposed in this work to extract features automatically, without manual interference, to detect the clinical acceptability of ECG signals efficiently. In order to create DCNN, the deformable convolution and pooling layers are merged into the regular convolutional neural network (CNN) architecture. In DCNN, the equidistant sampling locations of a regular CNN are replaced with adaptive sampling locations, which improves the network’s ability to learn based on the input. Deformable convolution layers concentrate more on significant segments of the ECG signals rather than giving equal attention to all segments. The proposed method is able to detect acceptable and unacceptable ECG signals with an accuracy of 99.50%, recall of 99.78%, specificity of 99.60%, precision of 99.47%, and F-score of 0.999. Experimental results show that the proposed method performs better than earlier state-of-the-art techniques.

An Accurate and Fast Animal Species Detection System for Embedded Devices

Encounters between humans and wildlife often lead to injuries, especially in remote wilderness regions, and highways. Therefore, animal detection is a vital safety and wildlife conservation component that can mitigate the negative impacts of these encounters. Deep learning techniques have achieved the best results compared to other object detection techniques; however, they require many computations and parameters. A lightweight animal species detection model based on YOLOv2 was proposed. It was designed as a proof of concept of and as a first step to build a real-time mitigation system with embedded devices. Multi-level features merging is employed by adding a new pass-through layer to improve the feature extraction ability and accuracy of YOLOv2. Moreover, the two repeated 3 × 3 convolutional layers in the seventh block of the YOLOv2 architecture are removed to reduce computational complexity, and thus increase detection speed without reducing accuracy. Animal species detection methods based on regular Convolutional Neural Networks (CNNs) have been widely applied; however, these methods are difficult to adapt to geometric variations of animals in images. Thus, a modified YOLOv2 with the addition of deformable convolutional layers (DCLs) was proposed to resolve this issue. Our experimental results show that the proposed model outperforms the original YOLOv2 by 5.0% in accuracy and 12.0% in speed. Furthermore, our analysis shows that the modified YOLOv2 model is more suitable for deployment than YOLOv3 and YOLOv4 on embedded devices.