Dip Nandi

BanglaCrack-10 is a real-world building crack and surface-condition image dataset collected in Bangladesh for computer-vision and automated structural-surface inspection research. The public YOLOv8 release contains 3,603 JPG images from selected sites in Bogura and Dhaka and 3,622 axis-aligned bounding-box annotations across ten classes: Diagonal Cracks, Hairline Cracks, Horizontal Cracks, Map Pattern Cracks, No Crack, Shrinkage Cracks, Spalling Cracks, Stair Step Cracks, Structural Overload Cracks, and Vertical Cracks. Source photographs were acquired using iQOO Neo 11 and vivo Y51 smartphones under natural environmental conditions and varying viewpoints and camera-to-surface distances. The released images are standardized to 640 × 640 pixels and are organized into 2,571 training, 691 validation, and 341 test images. The dataset is intended for reuse in object detection, crack classification, building-surface inspection, transfer learning, and structural-health-monitoring research. Manual screening identified no duplicate images in the released dataset. However, the current partition is not described as building-level or wall-level leakage-free because explicit physical-group identifiers are unavailable and group-level separation has not been independently verified.

Deep learning models have demonstrated remarkable potential in breast cancer detection, yet their vulnerability to adversarial attacks raises significant concerns for clinical deployment. This study proposes a novel hybrid adversarial distillation framework that integrates ensemble adversarial training with knowledge distillation to develop a robust yet computationally efficient breast cancer detection model. We first establish the extreme vulnerability of standard deep learning models to adversarial perturbations, demonstrating accuracy drops from 99.69% to near-zero under projected gradient descent (PGD) attacks. To address this, we develop an ensemble teacher model combining DenseNet121, ResNet50, and EfficientNet-B0 trained adversarially on FGSM-perturbed images. Knowledge from this robust teacher is then distilled into a lightweight ResNet18 student model through a specialized loss function balancing cross entropy and Kullback-Leibler divergence. Extensive evaluations show that our distilled model maintains near 60% accuracy under strong FGSM attacks and, remarkably, 58.50% accuracy against PGD attacks (ϵ = 0.3, α = 0.01, 10 iterations), where both baseline and ensemble models fail completely. Under stronger attack setting (50 iterations), the model continues to maintain meaningful accuracy, indicating robustness beyond the standard PGD configuration. Furthermore, the distilled model demonstrates superior resistance to transfer attacks, maintaining over 70% accuracy against examples crafted on the teacher model. While the trade-off reduces clean-data accuracy to 78.88%, our approach provides a computationally efficient solution for securing breast cancer detection systems against adversarial manipulation. This work represents a significant step toward clinically viable, secure AI-based diagnostic tools, though future work should focus on further improving clean-data accuracy while maintaining robust adversarial defenses.

EEG-based brain-computer interfaces remain constrained by signal non-stationarity, noise, and high dimensionality, limiting reliable decoding of motor imagery and cognitive states. EEG-DATNet is a hybrid architecture addressing these challenges through four components: multi-scale temporal convolutions spanning 60-260,ms capturing oscillatory content across canonical frequency bands; depthwise spatial convolutions encoding inter-electrode functional dependencies; a dual statistical pooling module pairing mean-activity and local-variance streams, the latter targeting event-related desynchronization patterns inaccessible to average pooling; and a spiking activation function trained via surrogate gradients yielding approximately 60% output sparsity, supporting energy-efficient inference. These streams feed a dual-attention transformer encoder applying independent temporal and spatial attention.On BCI Competition IV-2a (4-class motor imagery), EEG-DATNet achieves a mean accuracy of 82.84% (kappa = 0.77), reaching 96.82% (kappa = 0.95) for the best subject. On IV-2b (2-class), mean accuracy is 89.03%, outperforming EEG-Conformer (87.27%), EEGNet (85.04%), and FBCNet (81.95%). On the Thinking Out Loud dataset for inner speech decoding, EEG-DATNet achieves 43.68% using only 22 channels, versus 36.1% for BiLSTM and 29.67% for EEGNet. Ablation experiments confirm that dual attention is structurally necessary, with removal reducing accuracy by over 30 percentage points for certain subjects. Under Gaussian noise simulating motion artefacts, accuracy degrades by only 2.1%, versus 3.0% for EEGNet and 2.5% for EEG-Conformer. Preliminary estimates place inference at approximately 0.47, GFLOPs and 3.2 ms latency, and fine-tuning with 20% subject-specific data retains the majority of peak accuracy, suggesting practical viability in data-scarce clinical settings.

Skin cancer’s rising incidence necessitates precise and timely diagnosis to improve patient prognoses.While deep learning models offer significant advancements in medical image analysis, their inherent"black-box" nature fundamentally hinders clinical acceptance and integration. This paper introducesan interpretable hybrid deep learning framework for accurate skin lesion classification, addressingthis critical gap. Our novel architecture combines a sophisticated convolutional neural network (CNN)feature extractor with a powerful LightGBM classifier. The custom CNN integrates advanced modules,including YOLO-inspired blocks, Squeeze-and-Excitation (SE) mechanisms, Fused MBConv layers,and a unique Transformer-based self-attention and attention pooling module, designed to captureintricate and discriminative dermatoscopic features. Evaluated on an extensive and diverse datasetwith a combination of ISIC 2018, ISIC 2019, and HAM10000 challenges utilizing 5-fold stratifiedcross-validation, the hybrid model obtained an accuracy of 85.65% for binary(benign/malignant)classification and 94.84% accuracy for multiclass HAM10000 dataset. Crucially, we employ acomplementary suite of explainable AI (XAI) techniques: SHapley Additive exPlanations (SHAP)values interpret LightGBM’s decisions based on CNN features, while Saliency Maps and Gradient-weighted Class Activation Mapping (Grad-CAM) visually pinpoint critical image regions drivingCNN predictions. These multi-modal explainability insights provide unprecedented transparencyinto the model’s rationale, highlighting clinically relevant dermatological features and fosteringindispensable trust. This work thus bridges the abyss between advanced AI capabilities and practicalclinical utility, enabling more robust and reliable skin cancer diagnosis.

With the technological advancements, global communication has largely shifted to text-based communication. As a result, the process of extracting meaningful insights from human behavior by machine learning techniques applied to textual data has now been significantly simplified. This research utilizes text mining methods to analyze customers feedback from food reviews, employing them as effective tools for opinion analysis and rating prediction from feedback. This research utilizes two neural network techniques (Normal Neural Network and LSTM) to analyze textual data and generate predicted scores ranging from 1 to 5 for each review from Amazon food review dataset. After implementing two neural network models, the system automatically generates a predicted score on a scale from 1 to 5. This study employs widely-used neural network techniques and provides a foundation for advancing text-based emotion detection in future research. The primary focus of this study is on unaltered customer feedback and it aims to solve the problem of accurately analyzing customer sentiments or opinion and extracting meaningful insights from their feedback. By comparing the performance of LSTM and standard neural networks, we achieve a 62.12% accuracy, showcasing superior results in emotion prediction from unstructured textual reviews. These insights pave the way for more scalable and efficient solutions in text mining for emotion detection.