Pre-trained Deep Convolutional Neural Network Research Articles

Feature-level Fusion vs. Score-level Fusion for Image Retrieval Based on Pre-trained Deep Neural Networks

Today’s complex multimedia content made retrieving images similar to the user’s query from the database a challenging task. The performance of a Content-Based Image Retrieval System (CBIR) system highly depends on the image representation in a form of low-level features and similarity measurement. The traditional visual descriptors that do not provide good prior domain knowledge could lead to poor performance retrieval results. On the other hand, Deep Convolutional Neural Networks (DCNNs) have recently achieved a remarkable success as methods for image classification in various domains. Recently, pre-trained deep convolution neural networks on thousands of classes have the ability to extract very accurate and representative features which, in addition to classification, can also be successfully used in image retrieval systems. ResNet152, GoogLeNet and InceptionV3 are some of the effective and successful examples of pre-trained DCNNs recently applied in a computer vision tasks such as object recognition, clustering, and classification. In this paper, two approaches for a CBIR system, namely early fusion and late fusion, have been presented and compared. The early fusion utilizes concatenation of the features extracted by each possible pair of DCNNs, that is ResNet152-GoogLeNet, ResNet152-InceptionV3, and GoogLeNet-InceptionV3, and the late fusion apply CombSum method with Z-Score standardization to combine the score results provided by each DCNN of the aforementioned pairs. In the experiments on a popular WANG dataset it has been shown that late fusion approach slightly outperforms early fusion approach. The best performance of our experiments in terms of Average Precision (AP) for the top 20 results reaches 96.82%.

E-DFu-Net: An efficient deep convolutional neural network models for Diabetic Foot Ulcer classification.

The Diabetic Foot Ulcer (DFU) is a severe complication that affects approximately 33% of diabetes patients globally, often leading to limb amputation if not detected early. This study introduces an automated approach for identifying and classifying DFU using transfer learning. DFU is typically categorized into ischemic and infection states, which are challenging to distinguish visually. We evaluate the effectiveness of pre-trained Deep Convolutional Neural Network (DCNN) models for autonomous DFU detection. Seven models are compared: EfficientNetB0, DenseNet121, ResNet101, VGG16, MobileNetV2, InceptionV3, and InceptionResNetV2. Additionally, we propose E-DFu-Net, a novel model derived from existing architectures, designed to mitigate overfitting. Experimental results demonstrate that E-DFu-Net achieves remarkable performance, with 97% accuracy in ischemia classification and 92% in infection classification. This advancement enhances current methodologies and aids practitioners in effectively detecting DFU cases.

A single modality apparent first impression personality recognition model with temporal emotion based LSTM

Apparent first impression prediction has made great progress with deep neural networks. There is a trend for multimodal fusion where features from different sources are fused together to improve the accuracy of the prediction. However, in a real-life scenario, it is often hard to gather features from different sources such as audio and background information. It is desirable to develop a method that could improve the prediction accuracy from a single source rather than multiple sources. This study developed a method to predict personality traits from a single source of information, i.e., facial information. Specifically, a pre-trained Deep Convolutional Neural Network was employed to extract emotional expression frame by frame in the video clip, which was then fed into a Long Short Term Memory model to predict the “Big Five” personality traits score. In Parallel, the model based on the static apparent facial features was trained, and finally, the facial feature and facial expression were fused with demographic data (age and gender). The proposed system is tested on the CharLearn Dataset and achieved an accuracy score of 90.67% ranked just below the top 5 CharLearn Competition. The result also showed that the dynamic emotional pattern has a positive impact on first impression prediction, especially on extraversion.

Segmentation-Free Outcome Prediction from Head and Neck Cancer PET/CT Images: Deep Learning-Based Feature Extraction from Multi-Angle Maximum Intensity Projections (MA-MIPs).

We introduce an innovative, simple, effective segmentation-free approach for survival analysis of head and neck cancer (HNC) patients from PET/CT images. By harnessing deep learning-based feature extraction techniques and multi-angle maximum intensity projections (MA-MIPs) applied to Fluorodeoxyglucose Positron Emission Tomography (FDG-PET) images, our proposed method eliminates the need for manual segmentations of regions-of-interest (ROIs) such as primary tumors and involved lymph nodes. Instead, a state-of-the-art object detection model is trained utilizing the CT images to perform automatic cropping of the head and neck anatomical area, instead of only the lesions or involved lymph nodes on the PET volumes. A pre-trained deep convolutional neural network backbone is then utilized to extract deep features from MA-MIPs obtained from 72 multi-angel axial rotations of the cropped PET volumes. These deep features extracted from multiple projection views of the PET volumes are then aggregated and fused, and employed to perform recurrence-free survival analysis on a cohort of 489 HNC patients. The proposed approach outperforms the best performing method on the target dataset for the task of recurrence-free survival analysis. By circumventing the manual delineation of the malignancies on the FDG PET-CT images, our approach eliminates the dependency on subjective interpretations and highly enhances the reproducibility of the proposed survival analysis method. The code for this work is publicly released.

Diagnostic System Based on Deep Learning to Detect Diabetic Retinopathy

Purpose: To develop a machine learning based diabetic retinopathy screening system to help ophthalmologists for initial level screening. Study Design: Diagnostic accuracy study. Place and Duration of Study: Haldwani in a private hospital from January, 2023 to June, 2023. Methods: A total of 229 fundus images (people suffering from diabetic retinopathy)were used which had micro aneurysms, soft exudates, hard exudates and hemorrhages. We classified these images and pre-processed them by scaling, orienting, and color adjustments. With the help of various pre-processing techniques, we decreased the size of our dataset so that it can be handled efficiently by our model with optimal resources.Visual Geometry Group (VGG) is a type of pre-trained deep convolutional neural network (CNN). The term “deep” refers to the number of layers; the VGG-16 uses 16 and VGG-19 uses 19 convolutional layers respectively. The model was tested on fresh retinal dataset. Results: Our research has demonstrated promising results, achieving a high accuracy rate of 90% on a human dataset by utilizing VGG16 for feature extraction and a Logistic Regression classifier for classification. Conclusion: Ophthalmologists can utilize this machine learning based screening system for diabetic retinopathy screening.

Efficient Guided Grad-CAM Tuned Patch Neural Network for Accurate Anomaly Detection in Full Images

Deep learning-based anomaly detection in images has recently gained popularity as an investigative field with many global submissions. To simplify complex data analysis, researchers in the deep learning subfield of machine learning employ Artificial Neural Networks (ANNs) with many hidden layers. Finding data occurrences that significantly differ from generalizable to most data sets is the primary goal of anomaly detection. Many medical imaging applications use convolutional neural networks (CNNs) to examine anomalies automatically. While CNN structures are reliable feature extractors, they encounter challenges when simultaneously classifying and segmenting spots that need removal from scans. We suggest a separate and integration system to solve these issues, separated into two distinct departments: classification and segmentation. Initially, many network architecturesare taught independently for each abnormality, and these networks’ main components are combined. A sharedcomponent of the branched structure functions for all abnormalities. The final structure has two branches: onehas distinct sub-networks, each intended to classify a particular abnormality, and the other for segmenting various abnormalities. Deep CNNs training directly on high-resolution images necessitate input layer image compression, which results in the loss of information necessary for detecting medical abnormalities. A guided GradCAM (GCAM) tuned patch neural network is applied to full-size images for anomaly localization. Therefore, the suggested approach merges the pre-trained deep CNNs with class activation mappings and area suggestion systems to construct abnormality sensors and then fine-tunes the CNNs on picture patches, focusing on medical abnormalities instead of training on whole images. A mammogram data set was used to test the deep patch classifier, which had a 99% overall classification accuracy. A Brain tumor image data set was used to test the integrateddetector’s ability to detect abnormalities, and it did so with an average precision of 0.99.

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.

Using Advanced Convolutional Neural Network Approaches to Reveal Patient Age, Gender, and Weight Based on Tongue Images.

The human tongue has been long believed to be a window to provide important insights into a patient's health in medicine. The present study introduced a novel approach to predict patient age, gender, and weight inferences based on tongue images using pretrained deep convolutional neural networks (CNNs). Our results demonstrated that the deep CNN models (e.g., ResNeXt) trained on dorsal tongue images produced excellent results for age prediction with a Pearson correlation coefficient of 0.71 and a mean absolute error (MAE) of 8.5 years. We also obtained an excellent classification of gender, with a mean accuracy of 80% and an AUC (area under the receiver operating characteristic curve) of 88%. ResNeXt model also obtained a moderate level of accuracy for weight prediction, with a Pearson correlation coefficient of 0.39 and a MAE of 9.06 kg. These findings support our hypothesis that the human tongue contains crucial information about a patient. This study demonstrated the feasibility of using the pretrained deep CNNs along with a large tongue image dataset to develop computational models to predict patient medical conditions for noninvasive, convenient, and inexpensive patient health monitoring and diagnosis.

Enhanced PRIM recognition using PRI sound and deep learning techniques.

Pulse repetition interval modulation (PRIM) is integral to radar identification in modern electronic support measure (ESM) and electronic intelligence (ELINT) systems. Various distortions, including missing pulses, spurious pulses, unintended jitters, and noise from radar antenna scans, often hinder the accurate recognition of PRIM. This research introduces a novel three-stage approach for PRIM recognition, emphasizing the innovative use of PRI sound. A transfer learning-aided deep convolutional neural network (DCNN) is initially used for feature extraction. This is followed by an extreme learning machine (ELM) for real-time PRIM classification. Finally, a gray wolf optimizer (GWO) refines the network's robustness. To evaluate the proposed method, we develop a real experimental dataset consisting of sound of six common PRI patterns. We utilized eight pre-trained DCNN architectures for evaluation, with VGG16 and ResNet50V2 notably achieving recognition accuracies of 97.53% and 96.92%. Integrating ELM and GWO further optimized the accuracy rates to 98.80% and 97.58. This research advances radar identification by offering an enhanced method for PRIM recognition, emphasizing the potential of PRI sound to address real-world distortions in ESM and ELINT systems.

Automated algorithm selection using meta-learning and pre-trained deep convolution neural networks

Automated machine learning (AutoML), which explores the automation of various machine learning tasks, has seen significant progress in recent years. One of the fundamental tasks in this field is the selection of effective algorithms for a given dataset. This task is particularly challenging for tabular datasets where, unlike images or text, one cannot use pre-trained embeddings from other datasets. In this study, we present AutoIRAD, a novel meta-learning and vision-based approach for automatic classification algorithm selection for tabular datasets. Our approach is the first to generate image-based representations of entire tabular datasets, enabling us to model the features and interactions of all the dataset samples. Our image-based representation has two significant advantages. First, training our model across multiple datasets enables it to simultaneously learn from diverse domains and create more generalized and effective dataset representations. This representation can also be effectively applied to previously unseen datasets. Secondly, using images enables us to leverage large pre-trained networks (e.g., VGG). By tuning existing networks rather than training new ones, we can both speed up the training of our model and reduce the needed amounts of training data. An extensive evaluation of 150 datasets shows that AutoIRAD achieves results comparable to the leading state-of-the-art AutoML solution but at a fraction of their computation time.

Transfer Learning-based Drowsiness Detection System for Driver Assistance and Classification of Traffic Signs Employing a Deep Convolutional Neural Network

In this research paper, a comparative analysis of various image enhancements in the spatial domain techniques was performed based on three-dimensional image quality statistics such as mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity for measuring the image quality (SSIM). The conditions for choosing the best method are low mean square error (MSE) and high peak signal-to-noise ratio as well as the structural similarity for measuring the image quality. The pre-processed image was subjected to a classification task using a stepwise version of pre-trained deep convolutional neural networks. The problem of driver drowsiness caused by fatigue, lack of sleep, medication, etc. will have to be solved by improving the efficiency of driver drowsiness detection. Many traffic accidents are largely caused by drowsy drivers. Driver fatigue and distraction can cause a lack of alertness, which can lead to driver inattention. In this paper, we proposed a new deep-learning technique for driver drowsiness detection. Two methods using convolutional deep neural network and GoogleNet transfer learning are compared to achieve a better accuracy of 99%.

A fine-tuning deep residual convolutional neural network for emotion recognition based on frequency-channel matrices representation of one-dimensional electroencephalography

Emotion recognition (ER) plays a crucial role in enabling machines to perceive human emotional and psychological states, thus enhancing human-machine interaction. Recently, there has been a growing interest in ER based on electroencephalogram (EEG) signals. However, due to the noisy, nonlinear, and nonstationary properties of electroencephalography signals, developing an automatic and high-accuracy ER system is still a challenging task. In this study, a pretrained deep residual convolutional neural network model, including 17 convolutional layers and one fully connected layer with transfer learning technique in combination frequency-channel matrices (FCM) of two-dimensional data based on Welch power spectral density estimate from the one-dimensional EEG data has been proposed for improving the ER by automatically learning the underlying intrinsic features of multi-channel EEG data. The experiment result shows a mean accuracy of 93.61 ± 0.84%, a mean precision of 94.70 ± 0.60%, a mean sensitivity of 95.13 ± 1.02%, a mean specificity of 91.04 ± 1.02%, and a mean F1-score of 94.91 ± 0.68%, respectively using 5-fold cross-validation on the DEAP dataset. Meanwhile, to better explore and understand how the proposed model works, we noted that the ranking of clustering effect of FCM for the same category by employing the t-distributed stochastic neighbor embedding strategy is: softmax layer activation is the best, the middle convolutional layer activation is the second, and the early max pooling layer activation is the worst. These findings confirm the promising potential of combining deep learning approaches with transfer learning techniques and FCM for effective ER tasks.

Cervical Cancer Classification From Pap Smear Images Using Deep Convolutional Neural Network Models

As one of the most common female cancers, cervical cancer often develops years after a prolonged and reversible pre-cancerous stage. Traditional classification algorithms used for detection of cervical cancer often require cell segmentation and feature extraction techniques, while convolutional neural network (CNN) models demand a large dataset to mitigate over-fitting and poor generalization problems. To this end, this study aims to develop deep learning models for automated cervical cancer detection that do not rely on segmentation methods or custom features. Due to limited data availability, transfer learning was employed with pre-trained CNN models to directly operate on Pap smear images for a seven-class classification task. Thorough evaluation and comparison of 13 pre-trained deep CNN models were performed using the publicly available Herlev dataset and the Keras package in Google Collaboratory. In terms of accuracy and performance, DenseNet-201 is the best-performing model. The pre-trained CNN models studied in this paper produced good experimental results and required little computing time.Graphical

Diagnostic usefulness of deep learning methods for Helicobacter pylori infection using esophagogastroduodenoscopy images.

We aimed to assess the diagnostic potential of deep convolutional neural networks (DCNNs) for detecting Helicobacter pylori infection in patients who underwent esophagogastroduodenoscopy and Campylobacter-like organism tests. We categorized a total of 13,071 images of various gastric sub-areas and employed five pretrained DCNN architectures: ResNet-101, Xception, Inception-v3, InceptionResnet-v2, and DenseNet-201. Additionally, we created an ensemble model by combining the output probabilities of the best models. We used images of different sub-areas of the stomach for training and evaluated the performance of our models. The diagnostic metrics assessed included area under the curve (AUC), specificity, accuracy, positive predictive value, and negative predictive value. When training included images from all sub-areas of the stomach, our ensemble model demonstrated the highest AUC (0.867), with specificity at 78.44%, accuracy at 80.28%, positive predictive value at 82.66%, and negative predictive value at 77.37%. Significant differences were observed in AUC between the ensemble model and the individual DCNN models. When training utilized images from each sub-area separately, the AUC values for the antrum, cardia and fundus, lower body greater curvature and lesser curvature, and upper body greater curvature and lesser curvature regions were 0.842, 0.826, 0.718, and 0.858, respectively, when the ensemble model was used. Our study demonstrates that the DCNN model, designed for automated image analysis, holds promise for the evaluation and diagnosis of Helicobacter pylori infection.

Multimodal Gait Abnormality Recognition Using a Convolutional Neural Network-Bidirectional Long Short-Term Memory (CNN-BiLSTM) Network Based on Multi-Sensor Data Fusion.

Global aging leads to a surge in neurological diseases. Quantitative gait analysis for the early detection of neurological diseases can effectively reduce the impact of the diseases. Recently, extensive research has focused on gait-abnormality-recognition algorithms using a single type of portable sensor. However, these studies are limited by the sensor's type and the task specificity, constraining the widespread application of quantitative gait recognition. In this study, we propose a multimodal gait-abnormality-recognition framework based on a Convolutional Neural Network-Bidirectional Long Short-Term Memory (CNN-BiLSTM) network. The as-established framework effectively addresses the challenges arising from smooth data interference and lengthy time series by employing an adaptive sliding window technique. Then, we convert the time series into time-frequency plots to capture the characteristic variations in different abnormality gaits and achieve a unified representation of the multiple data types. This makes our signal processing method adaptable to several types of sensors. Additionally, we use a pre-trained Deep Convolutional Neural Network (DCNN) for feature extraction, and the consequently established CNN-BiLSTM network can achieve high-accuracy recognition by fusing and classifying the multi-sensor input data. To validate the proposed method, we conducted diversified experiments to recognize the gait abnormalities caused by different neuropathic diseases, such as amyotrophic lateral sclerosis (ALS), Parkinson's disease (PD), and Huntington's disease (HD). In the PDgait dataset, the framework achieved an accuracy of 98.89% in the classification of Parkinson's disease severity, surpassing DCLSTM's 96.71%. Moreover, the recognition accuracy of ALS, PD, and HD on the PDgait dataset was 100%, 96.97%, and 95.43% respectively, surpassing the majority of previously reported methods. These experimental results strongly demonstrate the potential of the proposed multimodal framework for gait abnormality identification. Due to the advantages of the framework, such as its suitability for different types of sensors and fewer training parameters, it is more suitable for gait monitoring in daily life and the customization of medical rehabilitation schedules, which will help more patients alleviate the harm caused by their diseases.

Object detection and classification of butterflies using efficient CNN and pre-trained deep convolutional neural networks

Object detection and classification of butterflies using efficient CNN and pre-trained deep convolutional neural networks

Standard fetal ultrasound plane classification based on stacked ensemble of deep learning models

Identifying standard fetal ultrasound (US) planes with key anatomical structures during mid-pregnancy prenatal screening is crucial for measuring fetal growth parameters and early detection of abnormalities. However, obtaining these standard planes is laborious and time-consuming and depends on the clinical experience of sonographers. Automatic detection of these planes can aid sonographers in identifying the correct standard planes. In recent times, various deep learning techniques have developed to automate the detection of standard fetal US planes. However, a common limitation among these approaches is their dependence on a single model prediction to make the final decision, which introduces the possibility of inaccuracies. Therefore, we propose an automated identification of commonly used standard fetal US planes based on the stacking ensemble of deep convolutional neural networks (CNN). The stacking ensemble method employs three pre-trained deep CNNs: AlexNet, VGG-19, and DarkNet-19. Softmax and random forest classifiers are used to get predictions from deep CNNs. The final prediction is made using the absolute majority voting technique. A publicly available fetal US dataset is employed to evaluate the performance of the stacking ensemble approach. The proposed ensemble model classifies fetal US planes into six distinct classes: abdomen, brain, femur, thorax, maternal cervix, and other (less commonly employed planes, such as kidney, and limbs) fetal planes. Experimental findings demonstrate that the stacking ensemble approach achieved high performance with an accuracy of 95.69%, precision of 94.02%, recall of 96.28%, F1-score of 95.08%, specificity of 99.12%, and Matthews correlation coefficient of 94.19% compared to individual deep CNN models and other competing methods.

Deep Learning Based Obstacle Awareness from Airborne Optical Sensors

Aviation statistics identify collision with terrain and obstacles as a leading cause of helicopter accidents. Assisting helicopter pilots in detecting the presence of obstacles can greatly mitigate the risk of collisions. However, only a limited number of helicopters in operation have an installed helicopter terrain awareness and warning system (HTAWS), while the cost of active obstacle warning systems remains prohibitive for many civil operators. In this work, we apply machine learning to automate obstacle detection and classification in combination with commercially available airborne optical sensors. While numerous techniques for learning-based object detection have been published in the literature, many of them are data and computation intensive. Our approach seeks to balance the detection and classification accuracy of the method with the size of the training data required and the runtime. Specifically, our approach combines the invariant feature extraction ability of pretrained deep convolutional neural networks (CNNs) and the high-speed training and classification ability of a novel, proprietary frequency-domain support vector machine (SVM) method. We describe our experimental setup comprising the CNN + SVM model and datasets of predefined classes of obstacles—pylons, chimneys, antennas, TV towers, wind turbines, helicopters—synthesized from prerecorded airborne video sequences of low-altitude helicopter flight. We analyze the detection performance using average precision, average recall, and runtime performance metrics on representative test data. Finally, we present a simple architecture for real-time, onboard implementation and discuss the obstacle detection performance of recently concluded flight tests.

Efficient pollen grain classification using pre-trained Convolutional Neural Networks: a comprehensive study

Pollen identification is necessary for several subfields of geology, ecology, and evolutionary biology. However, the existing methods for pollen identification are laborious, time-consuming, and require highly skilled scientists. Therefore, there is a pressing need for an automated and accurate system for pollen identification, which can be beneficial for both basic research and applied issues such as identifying airborne allergens. In this study, we propose a deep learning (DL) approach to classify pollen grains in the Great Basin Desert, Nevada, USA. Our dataset consisted of 10,000 images of 40 pollen species. To mitigate the limitations imposed by the small volume of our training dataset, we conducted an in-depth comparative analysis of numerous pre-trained Convolutional Neural Network (CNN) architectures utilizing transfer learning methodologies. Simultaneously, we developed and incorporated an innovative CNN model, serving to augment our exploration and optimization of data modeling strategies. We applied different architectures of well-known pre-trained deep CNN models, including AlexNet, VGG-16, MobileNet-V2, ResNet (18, 34, and 50, 101), ResNeSt (50, 101), SE-ResNeXt, and Vision Transformer (ViT), to uncover the most promising modeling approach for the classification of pollen grains in the Great Basin. To evaluate the performance of the pre-trained deep CNN models, we measured accuracy, precision, F1-Score, and recall. Our results showed that the ResNeSt-110 model achieved the best performance, with an accuracy of 97.24%, precision of 97.89%, F1-Score of 96.86%, and recall of 97.13%. Our results also revealed that transfer learning models can deliver better and faster image classification results compared to traditional CNN models built from scratch. The proposed method can potentially benefit various fields that rely on efficient pollen identification. This study demonstrates that DL approaches can improve the accuracy and efficiency of pollen identification, and it provides a foundation for further research in the field.

Deep Seasonal Network for Remote Sensing Imagery Classification of Multi-Temporal Sentinel-2 Data

As a medium-resolution multi-temporal data source, Sentinel-2 data has the potential to match the performance of using very-high-resolution (VHR) images in deep learning applications. To fully leverage the multi-temporal nature of Sentinel-2 data, we introduce the Deep Seasonal Network (DeepSN). This composite architecture combines a pre-trained deep convolutional neural network (DCNN) for visual feature extraction with a long short-term memory (LSTM) model to capture temporal information and make classification predictions. We evaluate the effectiveness of DeepSN on a Maasai Boma classification task in the Tanzania region. The DeepSN takes a sequence of four seasonal data, each spanning three months, for Boma prediction. Through cross-season validation experiments, we compare various advanced DCNNs and select EfficientNet as the backbone for DeepSN, as it performs the best. DeepSN with an EfficientNet backbone achieves a significant 19% improvement in the F1 score compared to plain EfficientNet for the Boma classification task. This work introduces a versatile composite architecture capable of handling multi-temporal data efficiently, providing flexibility in choosing the most suitable feature extraction backbone. The performance of DeepSN demonstrates the viability of utilizing medium-resolution multi-temporal data instead of high-resolution images for diverse tasks.