Explainable Deep Learning Framework for Binary Corrosion Image Classification Using Grad-CAM

Ensuring high-quality production of glass products is essential in modern manufacturing, where even minor defects such as scratches, cracks, or surface irregularities can significantly affect safety, performance, and usability. Traditional visual inspection methods are labour-intensive, time-consuming, and prone to human error, creating a strong need for automated and reliable defect detection systems. The main objective of this research is to develop an efficient Deep Learning (DL) based framework for accurate binary classification of defective and non-defective glass products while addressing limitations such as subtle defect visibility, variation in lighting conditions, and dataset imbalance. Furthermore, the study aims to evaluate model performance under different hyperparameter settings using an intelligent optimisation technique. The proposed work employs the Residual Network 50 Version 2 (ResNet50V2) Convolutional Neural Network (CNN) for defect classification, supported by a comprehensive preprocessing pipeline consisting of resizing, normalisation, greyscale conversion, and data augmentation. Bayesian Optimisation (BO) is integrated to fine-tune key hyperparameters, ensuring optimal learning efficiency and improved generalisation. The framework is trained using a curated glass defect dataset and evaluated using standard performance metrics to demonstrate its robustness and suitability for real-world industrial inspection scenarios. Experimental results show that the optimised ResNet50V2-BO model achieves 98.36% accuracy, along with a precision of 98%, recall of 93%, and F1-score of 95% for the defective class. The confusion matrix and PR curves further confirm the model’s ability to reliably distinguish subtle surface defects. These findings highlight the potential of the proposed approach as a fast, reliable, and scalable solution for automated quality control in glass manufacturing industries.

Aquaculture plays a crucial role in global food security; however, disease outbreaks remain a major constraint to sustainable production. Rapid and reliable detection of fish diseases is essential to reduce mortality, economic losses, and the misuse of antimicrobials in aquaculture systems. Conventional diagnostic approaches, such as clinical observation and bacterial culturing, are time-consuming, costly, and require specialized expertise. Recent advances in deep learning, particularly convolutional neural networks (CNNs), have shown promise in automating image-based disease detection. This study aimed to compare a lightweight three-layer CNN model with pre-trained deep learning architectures (VGG16, InceptionV3, and ResNet50) for multi-class classification of infected freshwater fish species using a balanced image dataset collected from aquaculture farms in Thailand. Images from clinically infected freshwater fish were collected during routine farm inspections across six provinces in Thailand. The dataset included 424 images of four species: Asian seabass (Lates calcarifer), red tilapia (Oreochromis sp.), snakeskin gourami (Trichopodus pectoralis), and snakeheads (Channa striata). After preprocessing, a balanced dataset of 56 images per class (totaling 224) was created. The dataset was divided into training (80%) and testing (20%) subsets. On-the-fly data augmentation techniques, such as rotation, brightness adjustment, flipping, shifting, shearing, and zooming, were applied to the training data to reduce overfitting. A lightweight, three-layer CNN model with stochastic gradient descent was used and compared with pre-trained architectures (VGG16, InceptionV3, and ResNet50). Model performance was assessed through accuracy, precision, recall, F1-score, confusion matrix analysis, and five-fold cross-validation. Among the evaluated models, InceptionV3 achieved the highest classification accuracy (56.82%), followed by VGG16 (43.18%) and the proposed CNN (38.64%), while ResNet50 performed poorly (25%). The InceptionV3 model also demonstrated higher average precision (63%), recall (57%), and F1-score (56.75%), indicating superior classification capabilities. Confusion matrix analysis revealed that InceptionV3 correctly classified 25 out of 44 test images, outperforming the proposed CNN (17 correct predictions) and VGG16 (19 correct predictions). Five-fold cross-validation further confirmed the stability and relatively better performance of the InceptionV3 model. The comparative evaluation shows that pre-trained CNN architectures, especially InceptionV3, outperform a lightweight three-layer CNN when trained on small, balanced datasets of infected fish images. Although the proposed lightweight CNN has limited accuracy, its low computational needs suggest it could be useful in resource-limited aquaculture settings. Incorporating deep learning-based image analysis into aquaculture health monitoring systems could enable quick disease triage, support timely management decisions, and promote better biosecurity and sustainable fish production. Future research should increase the dataset size, include more fish species and disease types, and test model performance across different farming environments to improve its generalizability and practical use.

The accurate identification of collision positions on containers is critical in logistics and trade for enhancing cargo safety and determining accident liability. Traditional visual inspection methods are labor-intensive, time-consuming, and costly. This study leverages data from an Inertial Measurement Unit sensor and evaluates combinations of machine learning models and feature selection methods to identify the optimal approach for collision position detection. Five machine learning models (decision tree, k-nearest neighbors, support vector machine, random forest, and extreme gradient boosting) and five feature selection methods (Pearson’s correlation coefficient, mutual information, sequential forward selection, sequential backward selection, and extra trees) were assessed using three performance metrics: accuracy, execution time, and CPU utilization. Statistical analysis with the Friedman test confirmed significant differences in model and feature selection performance. The combination of k-nearest neighbors and extra trees achieved the highest accuracy of 97.1%, demonstrating that inexpensive IMU acceleration data can provide a cost-effective, efficient, and reliable solution for collision detection. This has strong practical implications for improving accident accountability and reducing inspection costs in the logistics industry.

A sedentary lifestyle can lead to heart disease, cancer, and type 2 diabetes. An anaerobic exercise called pilates can address these problems. Although pilates training can provide health benefits, the heavy load of pilates poses may cause severe muscle injury if not done properly. Surveys have found that many teenagers are unaware of the movements in pilates poses. Therefore, a system is needed to help users classify pilates poses accurately. MediaPipe is a system that accurately extracts the real time human body skeleton. Convolutional Neural Network (CNN) with transfer learning is an accurate method for image classification. There have been several studies investigated pilates poses classification. However, there is still no research applies the MediaPipe as a skeleton feature extractor and CNN with a transfer learning to classify pilates poses. In addition, previous research still does not implement the pilates poses classification in real-time. Based on this problem, this study creates a system using MediaPipe as a feature extractor and CNN with transfer learning as a real-time pilates poses classifier. This system runs on a mobile device and gets information from a camera sensor. The results from MediaPipe then be classified by pre-trained CNN architectures with transfer learning: MobileNetV2, Xception, and ResNet50. The best model was obtained by MobileNetV2, which had an f1 score of 98%. Ten people who didn't know much about Pilates also tested the system. They all agreed that the app could accurately identify Pilates poses, make people more interested in Pilates, and help them learn more about Pilates.

Lung diseases such as COVID-19, pneumonia, and tuberculosis remain major public health challenges worldwide, emphasizing the urgent demand for accurate and efficient diagnostic methods. This research explores the use of a Convolutional Neural Network (CNN)-based framework for binary classification of chest X-ray images to detect abnormalities. The methodology incorporates preprocessing techniques such as image resizing, normalization, data augmentation, and grayscale transformation to improve input data quality. CNN architecture comprising convolutional, pooling, fully connected, and dropout layers were trained and evaluated on publicly available datasets. The model attained a test accuracy of 92%; nevertheless, performance metrics revealed a disparity between the two classified categories. Class 0 (Normal) had precision (83%) and recall (90%), resulting in an F1-score of 0.80, whereas Class 1 (Abnormal) demonstrated higher precision (88%) and recall (90%) with an F1-score of 0.88. This highlights the need for further optimization to enhance the detection of normal cases. The findings underscore the potential of CNNs in automating lung disease detection but also reveal areas for improvement in model robustness and class balance.

Pneumonia remains a significant public health challenge, particularly in resource-limited settings where access to expert radiological diagnosis is scarce. This study proposes a deep learning-based approach using a custom Convolutional Neural Network (CNN) for the binary classification of chest X-ray images into “Pneumonia” and “Normal” categories. The model was trained and evaluated on a curated dataset of 5,856 chest X-ray images, incorporating data preprocessing and augmentation techniques to enhance generalizability. Evaluation of the proposed CNN yielded strong performance metrics, including an accuracy of 96.05%, a precision of 98.79%, a recall of 95.76%, and an AUC of 0.9921. The precision-recall curve also demonstrated an average precision score of 0.9970, confirming the model’s robustness, even under class imbalance. These results highlight the potential of the proposed CNN model to assist clinicians in rapid and accurate pneumonia diagnosis, supporting its applicability in clinical and low-resource healthcare environments.

The adhesion quality is one of the critical factors affecting EIFS quality. The traditional visual inspection method is time-consuming and labor-intensive work because limited inspectors have to check all insulation units when they are installed. Thus, this study proposes a CV-based automatic adhesion quality-inspection model for EIFS that can quickly and quantitatively inspect the shape and amount of adhesive. In the case study, the proposed models accurately determined the adhesion quality with measurement accuracies of 90.58%, 95.28%, and 93.58% in ribbon width, dab diameter, and adhesive area, respectively, and a root mean-square error of 6.9 mm in dab distance. The results of this study can achieve automation of inspection for EIFS.

The hom-thong banana, being a high-value Thai export variety, is facing significant risk from disease outbreaks affecting crop yield and quality. Traditional visual inspection methods in detection of diseases are labor consuming, error-prone. This research addresses these limitations by developing a new artificial intelligence (AI)-based automatic disease detection system for the hom-thong banana industry on top of cutting-edge computer vision technology. The study employed deep learning object detection models, contrasting Roboflow, you only look once (YOLO)v11, and YOLOv12 architectures, which were trained on a large dataset of 2,576 images of Thai banana plantations. With systematic data augmentation techniques, the dataset was augmented to 6,184 images of seven types of disease under varied environmental conditions. The method entailed extensive preprocessing and evaluation of performance through precision, recall, and mean average precision (mAP) metrics. Outcomes indicated that YOLOv12 outperformed with 93.3% accuracy, 83.3% sensitivity, and 86.3% mAP@50 compared to standard inspection schemes. This research is applicable to Thailand's smart agriculture initiative by providing farmers with low-cost, accurate, and effective disease monitoring equipment. The application of this AI system has the ability to enhance the yield of crops, reduce losses, and enhance the competitiveness of Thai banana exports in the global market, in support of sustainable agricultural development.

The reproductive status in abalone (Haliotis spp.) can be difficult to discern using traditional visual inspection methods and may therefore rely on lethal histological analyses for a definitive determination. With endangered species of abalone, lethal histological examinations are strictly prohibited due to their imperiled status. This research investigated the effectiveness of using non-invasive ultrasound technology as an alternate means of identifying the reproductive status of abalone rather than through invasive biopsies or lethal histological methods. Twelve abalone were randomly selected from cultured red abalone broodstock cohort (n = 200) produced at The Cultured Abalone Farm (Goleta, CA) to initially test these methods. Following initial gonad ultrasound determination, a subset of another cohort of 5 year-old (n = 62) red abalone produced at the UC Davis Bodega Marine Laboratory were monitored using ultrasonography to detect seasonal changes in gonad size for 7 weeks. Following the 7-week assessment period, the full cohort of 5-year old abalone (n = 122) were then spawned and gonads were also evaluated using ultrasound technology before and immediately after spawning. Ultrasound imaging technology accurately differentiated between digestive and reproductive tissues, allowing for the characterization of an ultrasound gonad index score for rapid assessment. Mean gonad thickness determined via ultrasound imaging was categorized on a scale of 1–5, with an index of 1 being the lowest (gonad tissue thin or absent) and an index of 5 being the highest (gonad margin has increased and is compressing the digestive gland) for red abalone. Ultrasound imaging technology was successful in rendering a gonad index score that can be used to track gonad maturation over time in cultured or captive species. Tracking ultrasound gonad scores over time is a useful tool for improving abalone culture production, and increasing animal welfare by reducing handling stress associated with gonad assessments. Individual abalone that are empirically determined ready to be spawned can then be selected for broodstock. Non-invasive ultrasound technology has the potential for broad applications within abalone aquaculture to enhance both food and conservation aquaculture breeding programs.

Introduction: A form of malignant tumour known as lymphoma originated in lymphoid hematopoietic organs. Because the physical characteristics of the many lymphoma classes are similar, accurately diagnosing lymphomas is one of the most difficult tasks. Hence, an efficient classification of lymphoma plays a very important role in order to provide patients with prompt care. The purpose of this work is to evaluate the performance of pre-trained Convolutional Neural Networks (CNNs) in the multiclass categorization of lymphomas. Objectives: Classification of Non-Hodgkin lymphomas by adopting pre-trained CNN architectures like ResNet50, VGG16, InceptionV3 and DenseNet201 are adopted. Utilize several pre-processing techniques for denoising, rescaling, and enriching the input images, including gaussian filter, min-max normalisation, and data augmentation. Perform a detailed performance analysis of the proposed work with existing models. Methods: This research uses the different CNN architectures such as VGG16, DenseNet201, InceptionV3 to classify the lymphoma. In pre-processing, the gaussian filter is used to denoise and smoothen the images, min-max normalization is used to rescale the images and the data augmentation is used for solving the data imbalance issue. Transfer Learning and Fine-Tuning is done which improves the overall performance of the model. Results: This study makes use of the multi cancer dataset from Kaggle. The performance of these pre-trained CNN models is evaluated using accuracy, precision, recall, and the F-measure. Based on simulation findings, DenseNet201 outperforms VGG16 and InceptionV3 with an accuracy of 99.90%. Furthermore, FFNN-ResNet50 and HPC are two current studies that are used to compare ResNet50 and DenseNet201. ResNet50-DenseNet201 has a high accuracy of 99.90% compared to FFNN-ResNet50 and HPC.   Conclusions: Several CNN architectures, including VGG16, InceptionV3 and DenseNet201 are employed in this study to categorize lymphomas. Several NHL classifications, including FL, CLL, and MCL, are classified using the pre-trained CNN architecture. The gaussian filter, which aids in smoothing the pictures, is used to eliminate noise from the histopathology images. The pixel limits are then scaled using min-max normalization to increase pixel intensity, and data augmentation is employed to prevent data imbalance problems. Improved categorization is achieved by the ResNet50 by extracting multi-scale characteristics from the images. Based on the simulation findings, it is evident that DenseNet201, which incorporates ResNet50 features, outperforms VGG16 and InceptionV3 due to the intricate interactions between data that dense connectivity enables. Furthermore, ResNet50-DenseNet201 performs better than FFNN-ResNet50 and HPC. In comparison to FFNN-ResNet50 and HPC, ResNet50-DenseNet201 has a high accuracy of 99.90%.

In this study, few convolutional neural networks (CNN) have been trained with a transfer learning method to facilitate either binary classification of radiography images into COVID-19 infected and normal or ternary classification into normal, pneumonia, and COVID-19 infected. As the number of COVID-19 cases grow exponentially, the proposed solution can provide an early home based computer-aided diagnosis to ease the pressure on healthcare. The decision made by the model can advise a patient on whether it is critical to visit a doctor or not. In this paper, a CNN based transfer learning model was used to provide a superior precision in image classification. The neural network model was trained and tested using 1,183 radiography images to report the precision that can be attained in authentic conditions using three different CNNs. The accuracy of the model in classifying radiography images is 97.46% for ternary classification and 99.36% accuracy for binary classification using VGG-16 CNN architecture. In addition, the tested algorithm is also developed as a web application for detecting COVID-19 with Chest X-ray images and deployed in the cloud for public use.

Palmprint verification is one of the most significant and popular approaches for personal authentication due to its high accuracy and efficiency. Using deep region of interest (ROI) and feature extraction models for palmprint verification, a novel approach is proposed where convolutional neural networks (CNNs) along with transfer learning are exploited. The extracted palmprint ROIs are fed to the final verification system, which is composed of two modules. These modules are (i) a pre-trained CNN architecture as a feature extractor and (ii) a machine learning classifier. In order to evaluate our proposed model, we computed the intersection over union (IoU) metric for ROI extraction along with accuracy, receiver operating characteristic (ROC) curves, and equal error rate (EER) for the verification task.The experiments demonstrated that the ROI extraction module could significantly find the appropriate palmprint ROIs, and the verification results were crucially precise. This was verified by different databases and classification methods employed in our proposed model. In comparison with other existing approaches, our model was competitive with the state-of-the-art approaches that rely on the representation of hand-crafted descriptors. We achieved a IoU score of 93% and EER of 0.0125 using a support vector machine (SVM) classifier for the contact-based Hong Kong Polytechnic University Palmprint (HKPU) database. It is notable that all codes are open-source and can be accessed online.

Knee osteoarthritis (KOA) is a prevalent degenerative joint disease requiring accurate severity assessment for effective management. Manual grading of KOA severity from X-ray images using the Kellgren-Lawrence (KL) system is subjective and time-consuming. Deep learning, particularly transfer learning with Convolutional Neural Networks (CNNs), offers a promising automated alternative. This study evaluates the performance of six widely used pre-trained CNN architectures (VGG16, VGG19, InceptionV3, DenseNet121, MobileNetV2, and ResNet101) applied by transfer learning to classify the severity of KOA into five KL grades (0-4) using the publicly available "Knee Osteoarthritis Dataset with Severity Grading". Models were trained using standard preprocessing and data augmentation. Performance was evaluated based on test accuracy and computational efficiency. Results indicate that VGG16 achieved the highest test accuracy (68.90%), closely followed by VGG19 (67.63%), while other architectures performed in the mid-60s percentage range. A distinct trade-off between accuracy and efficiency was observed, with larger VGG models being most accurate but slowest, and MobileNetV2 offering the fastest training with slightly lower accuracy. While demonstrating the feasibility of transfer learning for 5-class KOA grading, the performance suggests that achieving higher accuracy requires further optimization, potentially through deeper fine-tuning, class balancing strategies, or hybrid approaches. This work establishes a comparative baseline for applying widely used pre-trained models to the challenging task of knee osteoarthritis severity classification.

Indonesia is a country that has thousands of plant types that can be used as traditional medicine. However, some people have not utilized this potential optimally due to the lack of knowledge about medicinal plants' types, benefits, and substances. Therefore, there is a need to develop an application that can identify medicinal plants that grow in Indonesia and provide information about the benefits and content of the substances contained in them. In this study, medicinal plants will be recognized using a mobile application from leaf images based on a pre-trained convolutional neural network (CNN) with a transfer learning technique. Three pre-trained CNN architectures, namely VGG-16, MobileNetV2, and DenseNet-121, are explored for medicinal plant recognition. Hyperparameter tuning is performed at the fully connected layer of all architectures with 20 possible modifications to find the best model. The experimental results on 24 types of medicinal plants show that the model based on MobileNetV2 achieves the best classification accuracy of 97.74%. The best model is obtained by modifying the fully connected layer of MobileNetV2 into three dense layers with the number of neurons 736, 448, and 928, respectively. After the application recognizes the types of medicinal plants, information about the benefits and substances contained in them is displayed to the user.

Early detection of lung cancer greatly improves survival rates. However, it is still the leading cause of cancer-related deaths around the world. Radiologists often spend a lot of time on assessments, and these can sometimes be inconsistent when they diagnose from CT images. With the increasing availability of deep learning, especially convolutional neural networks (CNNs), automated tools present a promising way to improve diagnostic speed and reliability. This study looks at the performance of four pre-trained CNN architectures: DenseNet-169, EfficientNet-B7, Inception-V3, and ResNet-152. These networks help classify cancerous and non-cancerous lung tissue from CT scans. This project used a dataset of 5,000 images from The Cancer Imaging Archive (TCIA). To address the limited number of malignant cases, balanced the classes and used data augmentation. After fine-tuning, DenseNet-169 achieved the highest accuracy at 95.6%, followed by EfficientNet-B7, Inception-V3, and ResNet-152, with their respective accuracies noted. These findings show that CNN-based models can provide high diagnostic accuracy. They also offer a steady, fast alternative to manual image reading, making them good candidates for use in clinical settings, especially where resources are limited.