The proposed framework provides reproducible reference results for evaluating supervised machine learning approaches to defective pixel detection while analyzing the influence of feature extraction window size on classification performance and reconstructed image quality.
Abstract
Detecting defective pixels in CMOS Image Sensors is a critical task for ensuring high-quality image acquisition, as even minor defects can significantly degrade performance in vision systems and downstream processing. This study evaluates the performance of several supervised machine learning algorithms for defective pixel detection using 300 grayscale images of 512×512 pixels from the publicly available TAMPERE17 dataset, in which dead and hot pixels were randomly introduced. The evaluated algorithms include k-Nearest Neighbors (kNN), Naïve Bayes, Decision Tree, Random Forest, Support Vector Machine (SVM), and Multilayer Perceptron. Performance was assessed using specificity, recall, precision, and the phi coefficient. A sensitivity analysis was additionally performed using neighborhood windows of 3×3, 5×5, and 7×7 pixels during feature extraction. The results showed that the neighborhood size has a noticeable impact on classification performance, with the best overall results obtained using an SVM classifier with an RBF kernel and a 3×3 feature extraction window. This configuration achieved a phi coefficient of 98.08%, together with a specificity of 99.00%, a recall of 99.00%, and a precision of 99.02%. Statistical analysis further indicated that SVM (RBF) and Multilayer Perceptron exhibited statistically comparable performance under the evaluated experimental conditions. The analysis further revealed that defective pixels are primarily characterized by highly local intensity variations, while larger neighborhoods do not necessarily improve classification performance. Compared with previously reported defective pixel detection methods, the proposed approach achieved highly competitive results and the highest phi coefficient among the evaluated methods. Additionally, the impact of defective pixel detection on image quality was assessed using a simple median filter correction stage and evaluated through Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM). The best-performing configuration achieved a PSNR of 40.96 dB and an SSIM of 98.68% after median filter correction, demonstrating substantial improvement over median filtering alone, although the resulting image quality remains below that of the best-performing state-of-the-art correction methods. The proposed framework provides reproducible reference results for evaluating supervised machine learning approaches to defective pixel detection while analyzing the influence of feature extraction window size on classification performance and reconstructed image quality.
In order to enable accurate and efficient rice quality evaluation through full-surface defect detection of brown rice, a detection system based on convolutional neural network (CNN) was developed. A dataset of images of five categories—unhulled, normal, broken, cracked, and insect-bitten brown rice—was collected. Three CNN models, YOLOv5s, YOLOv7, and Faster R-CNN, were evaluated and compared with traditional algorithms including support vector machine (SVM) and back propagation (BP) neural networks. Experimental results showed that CNN-based methods in the present database significantly outperformed traditional approaches, with the YOLOv5s model achieving the best comprehensive performance: 95.80% detection accuracy, 10.90 ms inference time per image, and 92 frames/s processing speed. An improved Rice-YOLOv5s algorithm was further proposed and validated through batch detection experiments, achieving an average recognition accuracy of 96.44% and a processing time of 9.2 ms per image, which is equivalent to approximately 108.7 FPS. This study demonstrates the feasibility of CNN-based brown rice defect detection, with future work directed toward lightweight deployment and multimodal fusion for production-line application.
Zhao You, Jianchun Yan, Hai Wei et al.· Foods· 0 citations
Road surface conditions play a crucial role in ensuring transportation comfort and safety. Conventional road inspection methods that rely on manual observation are often time-consuming, expensive, and prone to subjectivity. This study proposes an automated approach to classify road surface conditions using texture-based feature extraction and machine learning algorithms. A total of 802 road images were independently collected, representing three classes: good, fair, and damaged. The images were preprocessed through resizing, grayscale conversion, Contrast Limited Adaptive Histogram Equalization (CLAHE), and pixel normalization to improve image quality. Texture features were then extracted using Gray Level Co-occurrence Matrix (GLCM), including contrast, homogeneity, energy, and correlation. The extracted features were used as input to four classification algorithms: Support Vector Machine (SVM), k-Nearest Neighbor (KNN), Random Forest, and Naive Bayes. Experimental results show that KNN achieved the best performance with 96.27% accuracy, followed by SVM and Random Forest with comparable results. Naive Bayes performed the lowest due to its detrimental assumption of feature independence. These findings demonstrate that texture-based features combined with appropriate machine learning algorithms can effectively classify road surface conditions. This approach has strong potential for implementation in automated, real-time road monitoring systems, especially on devices with limited computing resources, contributing to more efficient and objective infrastructure management.
Naufal Alif Vivaldi, N. Puspita, Hilda Hilda Mujaddidah et al.· Journal of Enhanced Studies...· 0 citations
The automatic partitioning of handwritten signatures into clean and noisy is crucial to reliable authentication in documents. A new approach is presented here that integrates traditional handcrafted texture features such as Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP) and Gray-Level Co-occurrence Matrix (GLCM) with deep learning models. We have performed experiments on an offline signature dataset DocSign V2 comprising 2,400 grayscale signature images (50 users, equal number of clean and noisy signatures) created for this work. Eight conventional classifiers are tested on individual and combinations of features. The performance of a CNN built for the task and also four transfer learning models - EfficientNetB0, ResNet50, MobileNetV2, and DenseNet121 - are compared. The experiments show the best result of 97.5% was achieved with the HOG-LBP+AdaBoost combination. This is better than the best of deep learning models, MobileNetV2, which reached 95%. This demonstrates that, particularly for small datasets, carefully designed handcrafted features can be used to effectively classify clean and noisy signatures automatically. Moreover, the proposed system can also be used as a post-processing module to enhance signatures for other analysis or verification.
Thejashwini B L, H S Nagendraswamy, Rajashekara M et al.· International journal of com...· 0 citations
We propose an adaptive image enhancement method based on photometric statistics to improve object detection under adverse illumination conditions. Conventional image enhancement methods primarily target perceptual quality and may alter recognition-relevant features, potentially degrading detection performance. In contrast, the proposed method adaptively determines the enhancement intensity by combining a predefined domain-specific preset with a photometric risk score calculated from the photometric statistics of the input image to estimate the risk of enhancement-induced photometric risks. Based on these estimates, the enhancement intensity is adaptively controlled, and unnecessary transformation is conditionally bypassed. Experiments on the Berkeley DeepDrive 100K dataset using YOLOv11 demonstrate that the proposed method improves F1-score and mean Average Precision by 0.101 and 0.195, respectively, compared with unprocessed images under adverse conditions, including low-light environments. These results demonstrate that detector-oriented adaptive enhancement can improve robustness while reducing performance degradation caused by unnecessary or excessive image transformation.
Chae-yeong Kim, Soon-kak Kwon· Applied Sciences· 0 citations
Image Processing Using Machine Learning is an intelligent Artificial Intelligence-based system designed to analyze,
process, classify, and enhance digital images using Machine Learning and Deep Learning techniques. In today's digital era,
enormous amounts of image data are generated from medical imaging systems, surveillance cameras, social media platforms,
satellite systems, and industrial applications. Traditional image processing methods often depend on manual feature extraction
and predefined algorithms, which face limitations in terms of accuracy, scalability, automation, and real-time performance. This
project addresses these challenges by providing an intelligent and automated image processing framework that utilizes Artificial
Intelligence, Machine Learning, and Computer Vision techniques.The system processes images through multiple stages,
including image acquisition, preprocessing, feature extraction, classification, segmentation, and image enhancement to generate
meaningful and accurate outputs. It analyzes image characteristics such as color, texture, edges, and shapes to identify patterns
and make intelligent decisions. The proposed system integrates modern technologies including Artificial Intelligence, Machine
Learning algorithms, Deep Learning models, Convolutional Neural Networks (CNNs), Computer Vision techniques, image
enhancement methods, and database management systems to ensure high accuracy, efficiency, scalability, and real-time
processing capabilities.By combining intelligent image analysis with automated learning and enhancement mechanisms, the
Image Processing Using Machine Learning system aims to provide a reliable, efficient, and scalable solution suitable for
healthcare diagnostics, surveillance systems, multimedia applications, autonomous vehicles, satellite image analysis,
agricultural monitoring, and industrial quality inspection systems.
R. Rajesh, R. Aswini· International Journal for Re...· 0 citations
The digital transformation of the wood-processing industry increasingly relies on computer vision, artificial intelligence (AI), and embedded sensing technologies to automate timber quality assessment. This study presents the laboratory validation (TRL4) of a low-cost embedded computer-vision demonstrator for automated surface-defect detection in beech (Fagus sylvatica) sawn timber, developed as the first operational module of the SMARTWOOD-AI technology-transfer platform. The system integrates a Raspberry Pi 4 single-board computer with a Sony IMX500 intelligent camera and employs an interpretable computer-vision pipeline based on 21 handcrafted colour (HSV), texture (Gray-Level Co-occurrence Matrix and Local Binary Patterns), and edge-density (Canny) features classified using a Random Forest algorithm. A dataset comprising 24 beech boards (192 labelled image regions) was evaluated using a board-level train/validation partitioning strategy to prevent data leakage. The classifier achieved an average accuracy of 88.2% (±7.1%) during five-fold cross-validation and 82.5% accuracy on an independent validation set, with high sensitivity for defect detection (recall = 0.93, F1-score = 0.88). The results demonstrate the technical feasibility of the proposed embedded inspection architecture and establish a reproducible experimental baseline for future integration of deep-learning models, digital twins, and intelligent cutting optimisation within the SMARTWOOD-AI platform.
Sorin-Eugen Popa, R. Grigore· International Journal of Eng...· 0 citations