ResNet Architecture Research Articles

Desert/Forest Fire Detection Using Machine/Deep Learning Techniques

As climate change and human activity increase the likelihood of devastating wildfires, the need for early fire detection methods is inevitable. Although, it has been shown that deep learning and artificial intelligence can offer a solution to this problem, there is still a lot of room for improvement. In this research, two new deep learning approaches to fire detection are developed and investigated utilizing pre-trained ResNet-50 and Xception for feature extraction with a detailed comparison against support vector machine (SVM), ResNet-50, Xception, and MobileViT architectures. Each architecture was tuned utilizing hyperparameter searches and trials to seek ideal combinations for performance. To address the under-representation of desert features in the current fire detection datasets, we have created a new dataset. This novel dataset, Utah Desert Fire, was created using controlled fires and aerial imaging with a DJI Mini 3 Pro drone. The proposed modified ResNet-50 architecture achieved the best performance on the Utah Desert Fire dataset, reaching 100% detection accuracy. To further compare the proposed methods, the popular forest fire detection dataset, DeepFire, was deployed with resulting performance analyzed against most recent literature. Here, our proposed modified Xception model outperformed latest publications attaining 99.221% accuracy. The performance of the proposed solutions show an increase in classification accuracy which can be leveraged for the identification of both desert and forest fires.

Classification of coastal shrimp species using deep learning InceptionResNetV2 with data augmentation

Coastal areas that are rich in shrimp natural resources usually have a strong fishing industry. Shrimp consumption can contribute to the local economy. Several types of shrimp have higher utilization potential in the fishing industry and human consumption. Classification helps sort out shrimp species in coastal waters. Deep learning technology helps identify and separate shrimp species through deep learning with the Convolutional neural network. The seven classes observed were Vaname Shrimp, Tiger Shrimp, Jerbung Shrimp, Giant Prawns, Red Shrimp, Ronggeng Shrimp, and Prawns. The data used is 50 per class, so 350 images are used. Due to dataset limitations, the proposed contribution is an initial Resnet architecture with augmentation. Resnet helps perform in-depth training to save training computation time, and improvements add variety to limited amounts of data. The result is that the average accuracy of the model is 99.4%, with an average training computation time of 7 and a half minutes. The MSE misclassification value was 0.0143, RMSE was 0.1195, and MAE was 0.0086. Testing tests and testing data on the model only takes 1-2 seconds. Prediction accuracy ranges from 90.01%-99.8%.

A Convolutional Neural Network for Automated Detection of Cervical Ossification of the Posterior Longitudinal Ligament using Magnetic Resonance Imaging.

Retrospective cohort study. We aimed to develop and validate a convolutional neural network (CNN) model to distinguish between cervical ossification of posterior longitudinal ligament (OPLL) and multilevel degenerative spinal stenosis using Magnetic Resonance Imaging (MRI) and to compare the diagnostic ability with spine surgeons. Some artificial intelligence models have been applied in spinal image analysis and many of promising results were obtained; however, there was still no study attempted to develop a deep learning model in detecting cervical OPLL using MRI images. In this retrospective study, 272 cervical OPLL and 412 degenerative patients underwent surgical treatment were enrolled and divided into the training (513 cases) and test dataset (171 cases). CNN models applying ResNet architecture with 34, 50, and 101 layers of residual blocks were constructed and trained with the sagittal MRI images from the training dataset. To evaluate the performance of CNN, the receiver operating characteristic curves of 3 ResNet models were plotted and the area under the curve were calculated on the test dataset. The accuracy, sensitivity, and specificity of the diagnosis by the CNN were calculated and compared with 3 senior spine surgeons. The diagnostic accuracies of our ResNet34, ResNet50, and ResNet101 models were 92.98%, 95.32%, and 97.66%, respectively; the area under the curve of receiver operating characteristic curves of these models were 0.914, 0.942, and 0.971, respectively. The accuracies and specificities of ResNet50 and ResNet101 models were significantly higher than all spine surgeons; for the sensitivity, ResNet101 model achieved better values than that of the 2 surgeons. The performance of our ResNet model in differentiating cervical OPLL from degenerative spinal stenosis using MRI is promising, better results were achieved with more layers of residual blocks applied.

A refined ResNet18 architecture with Swish activation function for Diabetic Retinopathy classification

Automatic detection of Diabetic Retinopathy (DR) is critically important, as it is the primary reason of irreversible loss of vision in the economically active populations in the developed countries. Early detection of the onset of Diabetic Retinopathy can greatly benefit clinical treatment; although several different feature extraction methods have been proposed, the task of retinal image classification remains tedious even for trained clinicians. This paper emphasizes on Diabetic Retinopathy detection as well as the analysis of the different stages of DR, performed on fundus images using Deep Learning algorithms. Fundus images of the patient were provided as input to the developed model evaluated using the real-time dataset of the hospital. The proposed ResNet-18 architecture with swish function has achieved an accuracy of 93.51%, sensitivity of 93.42%, precision of 93.77% and F1-score of 93.59%. The paper concludes with a comparative study of Simple CNN, VGGNet-16, MobileNet-V2 and ResNet architectures and other state-of-art approaches, which highlights ResNet-18 with Swish as the most effective deep learning classifier model for DR detection.

Efficient physics-based learned reconstruction methods for real-time 3D near-field MIMO radar imaging

Near-field multiple-input multiple-output (MIMO) radar imaging systems have recently gained significant attention. These systems generally reconstruct the three-dimensional (3D) complex-valued reflectivity distribution of the scene using sparse measurements. Consequently, imaging quality highly relies on the image reconstruction approach. Existing analytical reconstruction approaches suffer from either high computational cost or low image quality. In this paper, we develop novel non-iterative deep learning-based reconstruction methods for real-time near-field MIMO imaging. The goal is to achieve high image quality with low computational cost at compressive settings. The developed approaches have two stages. In the first approach, physics-based initial stage performs adjoint operation to back-project the measurements to the image-space, and deep neural network (DNN)-based second stage converts the 3D backprojected measurements to a magnitude-only reflectivity image. Since scene reflectivities often have random phase, DNN processes directly the magnitude of the adjoint result. As DNN, 3D U-Net is used to jointly exploit range and cross-range correlations. To comparatively evaluate the significance of exploiting physics in a learning-based approach, two additional approaches that replace the physics-based first stage with fully connected layers are also developed as purely learning-based methods. The performance is also analyzed by changing the DNN architecture for the second stage to include complex-valued processing (instead of magnitude-only processing), 2D convolution kernels (instead of 3D), and ResNet architecture (instead of U-Net). Moreover, we develop a synthesizer to generate large-scale dataset for training the neural networks with 3D extended targets. We illustrate the performance through experimental data and extensive simulations. The results show the effectiveness of the developed physics-based learned reconstruction approach compared to commonly used approaches in terms of both runtime and image quality at highly compressive settings. Our source codes and dataset are made available at https://github.com/METU-SPACE-Lab/Efficient-Learned-3D-Near-Field-MIMO-Imaging upon publication to advance research in this field.

Deep Cross-Training: An Approach to Improve Deep Neural Network Classification on Mammographic Images

Breast cancer is the leading cause of death among women. In Brazil, the estimated risk corresponds to 61.61 new cases per 100,000 women per year. The strategy of screening for the disease through mammography favors the early detection of the illness. However, the identification of lesions in mammographic images is a complex process, with technical and biological limitations. To deal with that, deep learning techniques have been employed in Computer-Aided Diagnosis Systems. This work presents an approach for improving the accuracy of cancer screening by applying deep learning: Deep Cross-Training. The approach was evaluated by classifying mammographic images from the INbreast and CBIS-DDSM databases, developing binary and multiclass models based on the ResNet-18 architecture. Through the experiments, it was possible to verify that the DCTr approach increased the classification accuracy and precision by 39% and 36%, respectively. Confirming an average precision of 0.760 ± 0.188 in predicting the ROIs of mammograms.

Classification of Crops through Self-Supervised Decomposition for Transfer Learning

The 2S-DT (Self-Supervised Decomposition for Transfer Learning) model, created for crop categorization using remotely sensed data, is a unique method introduced in this paper. It deals with the difficulty of incorrectly identifying crops with comparable phenology patterns, a problem that frequently arises in agricultural remote sensing. Two datasets from Nanajangudu taluk in the Mysore district, which has a widely varied irrigated agriculture system, are used to assess the model. Using self-supervised learning, the 2S-DT model addresses the misclassification issue that frequently occurs when working with unlabeled classes, especially in high-resolution images. It uses class decomposition (CD) layer and a downstream learning approach. Using the model’s learning and the particulars of each geographical context, this layer improves the information’s arrangement. Our model architecture’s foundation is ResNet, a well-known deep learning framework. Each residual block in our ResNet architecture is made up of two 3x3 convolutional layers. Each convolutional layer is followed by batch normalization and Rectified Linear Unit (ReLU) activation functions, which improve the model’s capacity for learning. We utilized a 7x7 convolutional layer with 64 filters and a stride of 2 for Conv1 in ResNet18, resulting in an output size of 112x112x64. Conv2, which consists of Res2a and Res2b, generated an output with the dimensions 48x48x64. Conv3, which included Res3a and Res3b, produced an output with the dimensions 28x28x128. These architectural selections were made with our experimental needs in mind. The 2S-DT model’s newly added features make it easier to identify classes and update weights, improving the stability of the features’ spatial and spectral data. Extensive tests performed on two datasets show the model’s viability. Overall accuracy has improved significantly, with the 2S-DT model surpassing comparable models like TVSM, 3DCAE, and GAN Model by obtaining 95.65% accuracy for dataset 1 and 88.91% accuracy for dataset 2.

Efficient Skip Connections-Based Residual Network (ESRNet) for Brain Tumor Classification.

Brain tumors pose a complex and urgent challenge in medical diagnostics, requiring precise and timely classification due to their diverse characteristics and potentially life-threatening consequences. While existing deep learning (DL)-based brain tumor classification (BTC) models have shown significant progress, they encounter limitations like restricted depth, vanishing gradient issues, and difficulties in capturing intricate features. To address these challenges, this paper proposes an efficient skip connections-based residual network (ESRNet). leveraging the residual network (ResNet) with skip connections. ESRNet ensures smooth gradient flow during training, mitigating the vanishing gradient problem. Additionally, the ESRNet architecture includes multiple stages with increasing numbers of residual blocks for improved feature learning and pattern recognition. ESRNet utilizes residual blocks from the ResNet architecture, featuring skip connections that enable identity mapping. Through direct addition of the input tensor to the convolutional layer output within each block, skip connections preserve the gradient flow. This mechanism prevents vanishing gradients, ensuring effective information propagation across network layers during training. Furthermore, ESRNet integrates efficient downsampling techniques and stabilizing batch normalization layers, which collectively contribute to its robust and reliable performance. Extensive experimental results reveal that ESRNet significantly outperforms other approaches in terms of accuracy, sensitivity, specificity, F-score, and Kappa statistics, with median values of 99.62%, 99.68%, 99.89%, 99.47%, and 99.42%, respectively. Moreover, the achieved minimum performance metrics, including accuracy (99.34%), sensitivity (99.47%), specificity (99.79%), F-score (99.04%), and Kappa statistics (99.21%), underscore the exceptional effectiveness of ESRNet for BTC. Therefore, the proposed ESRNet showcases exceptional performance and efficiency in BTC, holding the potential to revolutionize clinical diagnosis and treatment planning.

Active learning for extracting surgomic features in robot-assisted minimally invasive esophagectomy: a prospective annotation study

BackgroundWith Surgomics, we aim for personalized prediction of the patient's surgical outcome using machine-learning (ML) on multimodal intraoperative data to extract surgomic features as surgical process characteristics. As high-quality annotations by medical experts are crucial, but still a bottleneck, we prospectively investigate active learning (AL) to reduce annotation effort and present automatic recognition of surgomic features.MethodsTo establish a process for development of surgomic features, ten video-based features related to bleeding, as highly relevant intraoperative complication, were chosen. They comprise the amount of blood and smoke in the surgical field, six instruments, and two anatomic structures. Annotation of selected frames from robot-assisted minimally invasive esophagectomies was performed by at least three independent medical experts. To test whether AL reduces annotation effort, we performed a prospective annotation study comparing AL with equidistant sampling (EQS) for frame selection. Multiple Bayesian ResNet18 architectures were trained on a multicentric dataset, consisting of 22 videos from two centers.ResultsIn total, 14,004 frames were tag annotated. A mean F1-score of 0.75 ± 0.16 was achieved for all features. The highest F1-score was achieved for the instruments (mean 0.80 ± 0.17). This result is also reflected in the inter-rater-agreement (1-rater-kappa > 0.82). Compared to EQS, AL showed better recognition results for the instruments with a significant difference in the McNemar test comparing correctness of predictions. Moreover, in contrast to EQS, AL selected more frames of the four less common instruments (1512 vs. 607 frames) and achieved higher F1-scores for common instruments while requiring less training frames.ConclusionWe presented ten surgomic features relevant for bleeding events in esophageal surgery automatically extracted from surgical video using ML. AL showed the potential to reduce annotation effort while keeping ML performance high for selected features. The source code and the trained models are published open source.Graphical abstract

Enhancing healthcare security in the digital era: Safeguarding medical images with lightweight cryptographic techniques in IoT healthcare applications

The Internet of Things has allowed for the connectivity of medical imaging devices to the healthcare sector's data backbone. This development, made possible by the IoT, will speed up the diagnosis and treatment phases of medical care. The increasing reliance on interconnected devices and cloud-based systems creates potential entry points for cyber-attacks and unauthorized access to sensitive medical data which not only jeopardize patient privacy but also have serious implications for patient safety and trust in healthcare systems. Medical images are critical assets in healthcare, and their confidentiality, integrity, and availability are paramount for accurate diagnoses, treatment planning, and patient care. This instigated research related to medical image security in IoT healthcare. For this, we investigated the usage of a cryptography-based network for image encryption and decryption, and its potential application to the safe transmission of medical images through deep learning. In the proposed work, we use a key learning network based on the ResNet-50 architecture to do the mapping between various image representations. Due to the incorporation of these “hidden properties” into the learning model, the encryption technique can be fine-tuned for each specific domain. As a first step in decryption, we use reconstructive networks to convert the encrypted image back into its “plaintext” form. Once the hidden entities have been uncovered, a Return on Investment (ROI) framework can be made available, and data mining can be simplified by drawing directly from the user's local information environment. Imaging tools for gauging therapy is a highly reliable by using the proposed system. We used two different kinds of publicly available datasets, and that helped us succeed in our mission. The comprehensive empirical setup and the findings of the security analysis suggest that the suggested method may give an unprecedented level of security and an unprecedentedly powerful outcome.

Optimized Screening of Glaucoma using Fundus Images and Deep Learning

Diabetic retinopathy, glaucoma, and age-related macular degeneration are among the leading causes of global visual loss. Early detection and diagnosis of these conditions are crucial to reduce vision loss and improve patient outcomes. In recent years, deep learning algorithms have shown great potential in automating the diagnosis and categorization of eye disorders using medical photos. For this purpose, the ResNet-50 architecture is employed in a deep learning-based strategy. The approach involves fine-tuning a pre-trained ResNet-50 model using over 5,000 retinal pictures from the ODIR dataset, covering ten different ocular diseases. To enhance the model's generalization performance and avoid overfitting, various data augmentation techniques are applied to the training data. The model successfully detects glaucoma-related ocular illnesses, including cataract, diabetic retinopathy, and healthy eyes. Performance evaluation using metrics like accuracy, precision, recall, and F1-score shows that the model achieved 92.60% accuracy, 93.54% precision, 91.60% recall, and an F1-score of 91.68%. These results indicate that the proposed strategy outperforms many state-of-the-art approaches in the detection and categorization of eye disorders. This success underscores the potential of deep learning-based methods in automated ocular illness identification, facilitating early diagnosis and timely treatment to ultimately improve patient outcomes.

An end-to-end multimodal 3D CNN framework with multi-level features for the prediction of mild cognitive impairment

In recent years, deep learning methods based on brain image have been used for the diagnosis of cognitive impairment-related disorders. With the development of neuroimaging techniques, multi-modality image such as structural magnetic resonance imaging (sMRI) and positron emission tomography (PET) reflect structural and functional information of the brain respectively, and provide more techniques for the diagnosis of cognitive impairment diseases. Combining these complementary image features can lead to more accurate diagnostic assessments compared to using a single modality. Therefore, how to effectively combine multi-modality image features to realize the diagnosis of cognitive impairment disease needs to be further explored. In this work, we propose an end-to-end multimodal 3D CNN framework based on ResNet architecture, which integrates multi-level features obtained under the role of attention mechanisms to better capture subtle differences among brain images, and achieves remarkable diagnostic performance through spatial pyramid pooling strategy and effective fusion of multi-modality features. In this process, we demonstrate that the multimodal framework is more effective by means of non-shared parameters for multi-modality features learning. Moreover, the visualized attention maps show that our model can focus on important brain regions relevant to disease diagnosis. The experimental results demonstrated that our method improved the diagnostic performance in AD diagnosis and MCI conversion prediction by 6.37 % and 3.51 % compared to the single modality, and it also outperformed some recent state-of-the-art multimodal methods. Especially in AD diagnosis achieved an average accuracy of 94.61 %, which provides a more feasible technology for diagnostic assessment of patients with AD.

Res-Net-VGG19: Improved tumor segmentation using MR images based on Res-Net architecture and efficient VGG gliomas grading

BackgroundThe determination of area tumor presents the chief challenge in brain tumor therapy and assessment. Without ionizing radiation, the medical Magnetic Resonance Imaging (MRI) tool has appeared as an essential diagnostic technique for brain cancers. Using 2D MRI images, manual segmentation of brain tumor size is a slow, error-prone task which the performance is extremely depends on operator's experience. In that respect, a consistent totally automated segmentation approach for the brain tumor detection is effectively needed to get a proficient dimension of the tumor size. ResultsIn this paper, an effusively computerized scheme for brain tumor detection is proposed by the use of deep convolutional networks. The proposed method was appraised on Brain Tumor Image Segmentation (BRATS 2020) datasets, including 1352 affected by brain tumor. ConclusionCross-validation technique has revealed that our process can attain talented segmentation competently reaching higher accuracy compared to other previous studies.

Image Augmentation for BreaKHis Medical Data using Convolutional Neural Networks

In applying Convolutional Neural Network (CNN) to computer vision tasks in the medical domain, it is necessary to have sufficient datasets to train models with high accuracy and good general ability in identifying important patterns in medical data. This overfitting is exacerbated by data imbalances, where some classes may have a smaller sample size than others, leading to biased predictive results. The purpose of this augmentation is to create variation in the training data, which in turn can help reduce overfitting and increase the ability of the model to generalize. Therefore, comparing augmentation techniques becomes essential to assess and understand the relative effectiveness of each method in addressing the challenges of overfitting and data imbalance in the medical domain. In the context of the research described, namely a comparative analysis of augmentation performance on CNN models using the ResNet101 architecture, a comparison of augmentation techniques such as Image Generator, SMOTE, and ADASYN provides insight into which technique is most suitable for improving model performance on limited medical data. By comparing these techniques' accuracy, recall, and overall performance results, research can identify the most effective and relevant techniques in addressing the challenges of complex medical datasets. This provides a valuable guide for developing better CNN models in the future and may encourage further research in developing more innovative augmentation methods suitable for the medical domain.

Breast Cancer Detection in Histopathology Images using ResNet101 Architecture

Cancer is a significant challenge in many fields, especially health and medicine. Breast cancer is among the most common and frequent cancers in women worldwide. Early detection of cancer is the main step for early treatment and increasing the chances of patient survival. As the convolutional neural network method has grown in popularity, breast cancer can be easily identified without the help of experts. Using BreaKHis histopathology data, this project will assess the efficacy of the CNN architecture ResNet101 for breast cancer image classification. The dataset is divided into two classes, namely 1146 malignant and 547 benign. The treatment of data preprocessing is considered. The implementation of data augmentation in the benign class to obtain data balance between the two classes and prevent overfitting. The BreaKHis dataset has noise and uneven color distribution. Approaches such as bilateral filtering, image enhancement, and color normalization were chosen to enhance image quality. Adding flatten, dense, and dropout layers to the ResNet101 architecture is applied to improve the model performance. Parameters were modified during the training stage to achieve optimal model performance. The Adam optimizer was used with a learning rate 0.0001 and a batch size of 32. Furthermore, the model was trained for 100 epochs. The accuracy, precision, recall, and f1-score results are 98.7%, 98.73%, 98.7%, and 98.7%, respectively. According to the results, the proposed ResNet101 model outperforms the standard technique as well as other architectures.

Prediction of pulp exposure before caries excavation using artificial intelligence: Deep learning-based image data versus standard dental radiographs

ObjectivesThe objective was to examine the effect of giving Artificial Intelligence (AI)-based radiographic information versus standard radiographic and clinical information to dental students on their pulp exposure prediction ability. Methods292 preoperative bitewing radiographs from patients previously treated were used. A multi-path neural network was implemented. The first path was a convolutional neural network (CNN) based on ResNet-50 architecture. The second path was a neural network trained on the distance between the pulp and lesion extracted from X-ray segmentations. Both paths merged and were followed by fully connected layers that predicted the probability of pulp exposure. A trial concerning the prediction of pulp exposure based on radiographic input and information on age and pain was conducted, involving 25 dental students. The data displayed was divided into 4 groups (G): GX-ray, GX-ray+clinical data, GX-ray+AI, GX-ray+clinical data+AI. ResultsThe results showed that AI surpassed the performance of students in all groups with an F1-score of 0.71 (P < 0.001). The students’ F1-score in GX-ray+AI and GX-ray+clinical data+AI with model prediction (0.61 and 0.61 respectively) was slightly higher than the F1-score in GX-ray and GX-ray+clinical data (0.58 and 0.59 respectively) with a borderline statistical significance of P = 0.054. ConclusionsAlthough the AI model had much better performance than all groups, the participants when given AI prediction, benefited only ‘slightly’. AI technology seems promising, but more explainable AI predictions along with a 'learning curve' are warranted.

Multimodal Deep Learning with Discriminant Descriptors for Offensive Memes Detection

A meme is a visual representation that illustrates a thought or concept. Memes are spreading steadily among people in this era of rapidly expanding social media platforms, and they are becoming increasingly popular forms of expression. In the domain of meme and emotion analysis, the detection of offensives is a crucial task. However, it can be difficult to identify and comprehend the underlying emotion of a meme because its content is multimodal. Additionally, there is a lack of memes datasets that address how offensive a meme is, and the existing ones in this context have a bias towards the dominant labels or categories, leading to an imbalanced training set. In this article, we present a descriptive balanced dataset to help detect the offensive nature of the meme’s content using a proposed multimodal deep learning model. Two deep semantic models, baseline BERT and hateXplain-BERT, are systematically combined with several deep ResNet architectures to estimate the severity of the offensive memes. This process is based on the Meme-Merge collection that we construct from two publicly available datasets. The experimental results demonstrate the model’s effectiveness in classifying offensive memes, achieving F1 scores of 0.7315 and 0.7140 for the baseline datasets and Meme-Merge, respectively. The proposed multimodal deep learning approach also outperformed the baseline model in three meme tasks: metaphor understanding, sentiment understanding, and intention detection.

A comprehensive approach for PV wind forecasting by using a hyperparameter tuned GCVCNN-MRNN deep learning model

Precise power forecasting is important in contemporary energy management systems, especially for optimizing the use of renewable resources. However, precise predictions of solar and wind power face challenges due to their dependence on various environmental factors. In this research, we introduce a novel hybrid model that merges Convolutional Neural Network (CNN) and ResNet architectures with a Modified Recurrent Neural Network (MRNN). This MRNN model incorporates several key layers, including Bidirectional Long Short-Term Memory (BiLSTM), LSTM, and Bidirectional weighted Gated Recurrent Unit (BiGRU) layers. to address the constraints of conventional estimation methods and achieve precise short-term Photovoltaic and wind power prediction. Hyperparameters of the proposed CNN-MRNN model are optimized using the Grid Search CV algorithm, resulting in improved learning rate and performance. Comparative analysis with GCVCNN-RNN, GCVCNN-BiGRU,CNN-BiLSTM, and CNN-LSTM models demonstrates the superior performance of the proposed technique, achieving a 20% low mean absolute error (MAE) and 30% of root mean square error (RMSE). Additionally, the GCVCNN-MRNN model outperforms existing techniques based on Nash Sutcliffe analysis (0.9852) for time series forecasting. The experimental results validate the usefulness of the proposed GCVCNN-MRNN in correctly predicting PV and wind power, the model consistently achieves lower MAE and RMSE, along with higher correlation coefficient (CC) and R2, indicating improved accuracy and stability. This research contributes to the advancement of clean energy and expands the existing body of knowledge in power forecasting.

Exploring ResNet-18 Estimation Design through Multiple Implementation Iterations and Techniques in Legacy Databases

In a rapidly evolving landscape where automated systems and database applications are increasingly crucial, there is a pressing need for precise and efficient object recognition methods. This study contributes to this burgeoning field by examining the ResNet-18 architecture, a proven deep learning model, in the context of fruit image classification. The research employs an elaborate experimental setup featuring a diverse fruit dataset that includes Rambutan, Mango, Santol, Mangosteen, and Guava. The efficacy of single versus multiple ResNet-18 models is compared, shedding light on their relative classification accuracy. A unique aspect of this study is the establishment of a 90% decision threshold, introduced to mitigate the risk of incorrect classification. Our statistical analysis reveals a significant performance advantage of multiple ResNet-18 models over single models, with an average improvement margin of 15%. This finding substantiates the study’s central hypothesis. The implemented 90% decision threshold is determined to play a pivotal role in augmenting the system’s overall accuracy by minimizing false positives. However, it’s worth noting that the increased computational complexity associated with deploying multiple models necessitates further scrutiny. In sum, this study provides a nuanced evaluation of single and multiple ResNet-18 models in the realm of fruit image classification, emphasizing their utility in practical, real-world applications. The research opens avenues for future exploration by refining these methodologies and investigating their applicability to broader object recognition tasks.

Research on Key Technology of Ship Re-Identification Based on the USV-UAV Collaboration

Distinguishing ship identities is critical in ensuring the safety and supervision of the marine agriculture and transportation industry. In this paper, we present a comprehensive investigation and validation of the progression of ship re-identification technology within a cooperative framework predominantly governed by UAVs. Our research revolves around the creation of a ship ReID dataset, the creation of a ship ReID dataset, the development of a feature extraction network, ranking optimization, and the establishment of a ship identity re-identification system built upon the collaboration of unmanned surface vehicles (USVs) and unmanned aerial vehicles (UAVs). We introduce a ship ReID dataset named VesselID-700, comprising 56,069 images covering seven classes of typical ships. We also simulated the multi-angle acquisition state of UAVs to categorize the ship orientations within this dataset. To address the challenge of distinguishing between ships with small inter-class differences and large intra-class variations, we propose a fine-grained feature extraction network called FGFN. FGFN enhances the ResNet architecture with a self-attentive mechanism and generalized mean pooling. We also introduce a multi-task loss function that combines classification and triplet loss, incorporating hard sample mining. Ablation experiments on the VesselID-700 dataset demonstrate that the FGFN network achieves outstanding performance, with a Rank-1 accuracy of 89.78% and mAP of 65.72% at a state-of-the-art level. Generalization experiments on pedestrian and vehicle ReID datasets reveal that FGFN excels in recognizing other rigid body targets and diverse viewpoints. Furthermore, to further enhance the advantages of UAV-USV synergy in ship ReID performance, we propose a ranking optimization method based on the homologous fusion of multi-angle UAVs and heterologous fusion of USV-UAV collaborative architecture. This optimization leads to a significant 3% improvement in Rank-1 performance, accompanied by a 73% reduction in retrieval time cost.