Mohammad Nishat Akhtar

Automated corrosion classification is crucial for industrial inspection. However, existing methods face severe class imbalance and strict computational constraints on edge devices. This study proposes a stepwise self-training framework for six-category corrosion classification. Leveraging a Vision Transformer and XGBoost classifier, this study proposes a class-aware proportional screening strategy that expands a small, labelled dataset with 19,964 unlabeled images to create a more balanced training set. Progressive training and knowledge distillation are then integrated to train a lightweight MobileNetV2 student model. The optimal fine-tuned model achieves a peak classification accuracy of 87.52%, improving upon the baseline MobileNetV2 by over 5%. Grad-CAM analysis confirms accurate focus on corroded regions, while mask-based fine-tuning demonstrates robustness, maintaining over 83% accuracy under 10% visual interference. This approach minimizes computational overhead without sacrificing accuracy, thereby offering a practical solution for real-world monitoring on resource-constrained devices.

Climate change, workforce shortages, and sustainability requirements create serious obstacles for oil palm farms, which contribute $65 billion annually. Automated monitoring solutions are very important because traditional manual tree counting techniques, which are widely used in the sector, have counting error rates of 15-25% and require significant human resources. In this comprehensive review, deep learning applications, specifically, convolutional neural networks (CNNs) for oil palm tree detection and counting are evaluated. Performance, constraints, and realistic deployment pathways are examined. Publications from 2016 to 2025 that focused on oil palm recognition using deep learning with quantitative measurements were found using a literature search across Scopus, Web of Science, IEEE Xplore, and Google Scholar. Architectures, dataset properties, and performance metrics were recorded using data extraction. Malaysia is at the forefront of cooperative networks that span 22 nations, according to an analysis of 47 datasets. With mature trees, modern CNN architectures improved with YOLO frameworks achieve >95% detection accuracy; nevertheless, for young trees, they show notable degradation (87.2% vs. 96.8% mAP). Cross-regional generalization (21.9 percentage point accuracy degradation), processing demands (450-650 ms inference), and financial obstacles are important obstacles. Real-time viability is demonstrated by edge-optimized models, which achieve 98.6\% accuracy with 80 ms inference. Geographic bias (68% Malaysian, 23% Indonesian) and restricted public availability (8%) are revealed by dataset analysis. Deep learning can significantly improve oil palm management by 15-20% compared to conventional techniques. Widespread adoption requires standardized benchmark datasets (10,000+ images), transfer learning techniques (<500 images per region), edge-optimized architectures (<100 ms inference), and phased deployment (10-50 hectare pilots). CNN's convergence with precision agriculture positions the industry for comprehensive digitalization while addressing sustainability and labor challenges.

Considering that corrosion is a widespread problem in tropical countries, this study proposes a progressive optimization of EfficientNetV2 for images of corroded objects (corrosion dataset) that can effectively target small and medium-sized corrosion datasets for detection. Compared to other models, the proposed model first adopts EfficientNetV2 as the basic architecture, focusing on the use of only MBConv blocks and MBConv with Fused-MBConv blocks in the hidden layers, as well as the effect of the number of these layers on the model’s classification results. To further improve the performance, this paper attempts to replace the convolutional modules in the input layer with LazyConv, utilizing FReLU and Dy-ReLU as an activation functions in both the input and output layers. The simulation results show that for the medium-sized corrosion dataset in this paper which uses only MBConv blocks for EfficientNetV2 can achieve higher accuracy but lower computational efficiency. Setting a smaller number of layers and replacing the convolutional block in the input layer with LazyConv can significantly reduce the total size of the model and make it more flexible, where the total size of the obtained M2 model being only 58.98 MB, and capable of automatically determining the number of input channels. Using FReLU in the input and output layers can achieve greater stability, with standard deviations of F1-score and Accuracy under five cycles of only 0.0099 and 0.0126, respectively. In addition, the optimized M2 model also offers advantages in terms of both light weight and stability compared to other classic deep learning models. These findings from this study may serve as a foundation for future innovations in the design of corrosion classification models.

This study investigates the application of YOLOv8 object detection models for identifying and counting oil palm trees in plantation management, with a focus on deployment on edge devices. Various YOLOv8 architectures were trained and evaluated using drone-captured images of palm oil plantations. The YOLOv8 nano model delivered superior performance, achieving 95.6% precision, 93.3% recall, and 98% mAP @0.5. The optimized model was deployed on a Raspberry Pi 4B (RPi) equipped with an Intel Neural Compute Stick 1 (NCS1) accelerator, enabling real-time detection under field conditions. Post-deployment performance analysis indicated an average inference time of 0.4 s per image (2.4 FPS) with minimal accuracy reduction (97% mAP @0.5). The system demonstrated low power consumption (1.6W peak) and efficient memory usage (450 MB RAM), emphasizing its suitability for edge computing in precision agriculture. This research demonstrates the feasibility of employing compact, efficient deep-learning models on resource-constrained devices like the RPi for palm oil tree detection and counting. The proposed system provides a portable and energy-efficient solution for real-time monitoring of palm oil plantations, offering significant potential to enhance plantation management by enabling data analysis and decision-making in remote or challenging environments. Despite promising results, the system's performance was evaluated under specific environmental conditions, potentially limiting its generalizability across diverse plantation landscapes. Future research should aim to expand detection capabilities to include disease identification and yield estimation. Graphical abstract

Occurrence of sudden expansion is widespread in the defense and automobile industry. At the blunt base of the fuselage, missiles, projectiles, and aircraft bombs, the flow gets separated at the base and forms low-pressure recirculation, leading to a significant increase in the base drag. This paper addresses how this low base pressure at the base can be controlled. A detailed numerical study was conducted to assess the impact of the quarter circle as a passive control mechanism in a suddenly expanded flow for an area ratio of 4.84 at Mach M = 2.0 for various rib radii ranging from 1 mm to 4 mm, the different duct lengths from L = 1D to 6D at different nozzle pressure ratio ranging from 3 to 11. The findings of this study show that a 1 mm rib is inadequate to impact the flow field inside the duct. Passive control in the form of a quarter circle rib seems to become effective once the nozzle flows under a favorable pressure gradient. However, the remaining rib radii effectively reduce the suction created at the base of the recirculation zone. It is found that the rib radius of 4 mm when placed at 66 mm at the base, results in a maximum rise in the base pressure, and the base pressure ratio attains a value of 3.4. These results are case sensitive; hence, one has to make a final decision about the rib radius, rib location, and the level of expansion based on the end user requirement.

This study assesses the effectiveness of passive control as a quarter circle at Mach M = 1 for a duct of diameter 22 mm of length L = 1D to 6D. The study was conducted for a nozzle pressure ratio of 1.5 to 5. At sonic Mach M = 1, when the passive control is placed at various locations in the duct, the optimum location and radius of the rib seem to be 1D, 1.5D, 2D, and 4 mm, resulting in a base pressure almost double the ambient pressure for orientation1 of the rib where flow from the Nozzle sees curved part of the rib. The base pressure equals the ambient pressure for a 3 mm rib radius at 1D, 1.5D, and 2D. When the rib is located at 0.5D, there is an increase in the base pressure due to the presence of the rib. However, the base pressure magnitude is lower than attained for 1D, 1.5D, 2D, and 3D rib locations. The best options for orientation 2 of the rib locations are 1.5D and 2D. No appreciable results are obtained when the rib is placed at 3D for both orientations.

The turbulent flow in a separated region is still a fundamental area of research due to the advent of space shuttles and high-performance military aircraft, and turbulent flow in transonic and supersonic flow is a thrust area for researchers. Whenever the flow experiences an abrupt increase in the area of the enlarged duct, the flow gets significant relief to separate and expand. When the shear layer comes out, it gets divided into two regions: main flow and separated flow. The divided stream line gets reattached with the duct and forms a recirculation zone where the pressure is lower than ambient pressure, resulting in significant drag. This study focuses on base pressure control through quarter-circle rib as a passive control mechanism. Accordingly, a comprehensive numerical simulation was carried out at screech-prone Mach number M = 1.6 for various radii 1 mm, 2 mm, 3 mm, and 4 mm for duct lengths in the range from L = 1D to 6D and nozzle pressure ratios from 3 to 11. Results indicate that for the same range of the rib radius, duct lengths, and level of expansion, there is a progressive increase in the base pressure when rib locations are moved downstream from 0.5D to 3D. The maximum rise in the base pressure is achieved when the rib is located at 66 mm from the base region. A rib with a radius of 1 mm is inadequate for the entire range of rib placement in the present study except when the rib is 11 mm from the base. It can be concluded that a rib of a 1.5 mm radius will be sufficient to neutralize the suction created due to the flow separation. The user can decide on the rib dimension, location, and nozzle pressure ratio based on their requirements.

The study of base pressure and its control is an important research area in the transonic speed when the flow undergoes a sudden change in area. The turbulent flow in a separated region is still a crucial area of research due to the advent of space shuttles and high-performance military aircraft, and turbulent flow in transonic and supersonic flow is a thrust area for researchers. This paper focuses on base pressure control with sudden expansion at Mach 1.3 for an area ratio of 4.84. The flow field inside the duct is controlled through a passive control in the form of quarter ribs of radii 1 mm, 2 mm, 3 mm, and 4 mm for various duct lengths in the range from L = 1D to 6D for nozzle pressure ratios in the range from 3 to 11. Results show that a 1 mm rib is not adequate, and rib radii 2 mm, 3 mm, and 4 mm are effective in raising the base pressure values, and this rise in the base pressure continues till the duct length L = 1D to 4D. There is a marginal reduction in base pressure for the duct lengths L = 5D and 6D due to the ineffectiveness of the back pressure.