YOLO-v1 to YOLO-v8, the Rise of YOLO and Its Complementary Nature toward Digital Manufacturing and Industrial Defect Detection

Hussain M., Al-Aqrabi H., Munawar M., Hill R., Parkinson S.

This paper presents a framework for the automated detection of Exudates, an early sign of Diabetic Retinopathy. The paper introduces a classification-extraction-superimposition (CES) mechanism for enabling the generation of representative exudate samples based on limited open-source samples. The paper demonstrates how the manipulation of Yolov5M output vector can be utilized for exudate extraction and super-imposition, segueing into the development of a custom CNN architecture focused on exudate classification in retinal based fundus images. The performance of the proposed architecture is compared with various state-of-the-art image classification architectures on a wide range of metrics, including the simulation of post deployment inference statistics. A self-label mechanism is presented, endorsing the high performance of the developed architecture, achieving 100% on the test dataset.

Lin G., Liu K., Xia X., Yan R.

Limited by computing resources of embedded devices, there are problems in the field of fabric defect detection, including small defect size, extremely unbalanced aspect ratio of defect size, and slow detection speed. To address these problems, a sliding window multihead self-attention mechanism is proposed for the detection of small targets, and the Swin Transformer module is introduced to replace the main module in the original YOLOv5 algorithm. First, to reduce the distance between several scales, the weighted bidirectional feature network is employed on embedded devices. In addition, it is helpful to improve the perception of small-target faults by incorporating a detection layer to achieve four-scale detection. At last, to improve the learning of positive sample instances and lower the missed detection rate, the generalized focal loss function is finally implemented on YOLOv5. Experimental results show that the accuracy of the improved algorithm on the fabric dataset reaches 85.6%, and the mAP is increased by 4.2% to 76.5%, which meets the requirements for real-time detection on embedded devices.

Binomairah A., Abdullah A., Khoo B.E., Mahdavipour Z., Teo T.W., Mohd Noor N.S., Abdullah M.Z.

Two common defects encountered during manufacturing of crystalline silicon solar cells are microcrack and dark spot or dark region. The microcrack in particular is a major threat to module performance since it is responsible for most PV failures and other types of damage in the field. On the other hand, dark region in which one cell or part of the cell appears darker under UV illumination is mainly responsible for PV reduced efficiency, and eventually lost of performance. Therefore, one key challenge for solar cell manufacturers is to remove defective cells from further processing. Recently, few researchers have investigated deep learning as an alternative approach for defect detection in solar cell manufacturing. The results are quite encouraging. This paper evaluates the convolutional neural network based on heavy-weighted You Only Look Once (YOLO) version 4 or YOLOv4 and the tiny version of this algorithm referred here as Tiny-YOLOv4. Experimental results suggest that the multi-class YOLOv4 is the best model in term of mean average precision (mAP) and prediction time, averaging at 98.8% and 62.9 ms respectively. Meanwhile an improved Tiny-YOLOv4 with Spatial Pyramid Pooling scheme resulted in mAP of 91.0% and runtime of 28.2 ms. Even though the tiny-weighted YOLOv4 performs slightly lower compared to its heavy-weighted counterpart, however the runtime of the former is 2.2 order much faster than the later.

Hussain M., Al-Aqrabi H., Hill R.

Photovoltaic cell manufacturing is a rigorous process involving many stages where the cell surface is exposed to external pressure and temperature differentials. This provides fertile ground for micro-cracks to develop on the cell surface. At present, domain experts carry out a manual inspection of the cell surface to judge if any micro-cracks are present. This research looks to overcome the issue of cell data scarcity through the proposed filter-induced augmentations, thus providing developers with an effective, cost-free mechanism for generating representative data samples. Due to the abstract nature of the cell surfaces, the proposed augmentation strategy is effective in generating representative samples for better generalization. Furthermore, a custom architecture is developed that is computationally lightweight compared to state-of-the-art architectures, containing only 7.01 million learnable parameters while achieving an F1-score of 97%.

Shi J., Yang J., Zhang Y.

Due to the irresistible factors of material properties and processing technology in the steel production, there may be different types of defects on the steel surface, such as rolling scale, patches and so on, which seriously affect the quality of steel, and thus have a negative impact on the economic efficiency of the enterprises. Different from the general target detection tasks, the defect detection tasks have small targets and extreme aspect ratio targets. The contradiction of high positioning accuracy for targets and their inconspicuous features makes the defect detection tasks difficult. Therefore, the original YOLOv5 algorithm was improved in this paper to enhance the accuracy and efficiency of detecting defects on steel surfaces. Firstly, an attention mechanism module was added in the process of transmitting the shallow feature map from the backbone structure to the neck structure, aiming at improving the algorithm attention to small targets information in the feature map and suppressing the influence of irrelevant information on the algorithm, so as to improve the detection accuracy of the algorithm for small targets. Secondly, in order to improve the algorithm effectiveness in detecting extreme aspect ratio targets, K-means algorithm was used to cluster and analyze the marked steel surface defect dataset, so that the anchor boxes can be adapted to all types of sizes, especially for extreme aspect ratio defects. The experimental results showed that the improved algorithms were better than the original YOLOv5 algorithm in terms of the average precision and the mean average precision. The mean average precision, demonstrating the largest increase among the improved YOLOv5 algorithms, was increased by 4.57% in the YOLOv5+CBAM algorithm. In particular, the YOLOv5+CBAM algorithm had a significant increase in the average precision for small targets and extreme aspect ratio targets. Therefore, the YOLOv5+CBAM algorithm could make the accurate localization and classification of steel surface defects, which can provide a reference for the automatic detection of steel defects.

Sun T., Xing H., Cao S., Zhang Y., Fan S., Liu P.

Accurate classification and detection of hot spots of photovoltaic (PV) panels can help guide operation and maintenance decisions, improve the power generation efficiency of the PV system, and ensure power stations’ safe and stable operation. Considering that, in this paper, the hot spots of PV panels collected on site are taken as the research object, and their formation mechanism is studied. Based on this, the morphological characteristics possessed by the hot spots of PV panels are classified into circular, linear, and array ones. A novel method for detecting hot spots of PV panels based on improved anchors and prediction heads of the YOLOv5 (AP-YOLOv5) network is proposed. Besides, to improve the detection precision of the YOLOv5 network at different scales in hot spots of PV panels, the K-means clustering algorithm is employed to cluster the length–width ratio of the data annotation frame, and a group of the anchors with smaller values is added so as to realize the detection of small targets by optimizing the cluster number. Apart from that, the corresponding prediction heads are constructed for the new anchor parameters to improve the detection precision concerning hot spots of PV panels. Furthermore, the model is verified by training experimental data and comparing test set results. The results showed that compared with other one-stage object detection models, the mean average precision (mAP) of the proposed network can achieve 87.8%, while the average recall rate is 89.0%, and the F1 score reaches 88.9%. In addition, the precision of this model is better than that of other models while maintaining a high frame rate; the frames per second (FPS) is as high as 98.6, thus laying a foundation for developing rapid detection tool of hot spots of PV panels, improving the safety of power station operation, and providing a method for intelligent operation and maintenance of PV power stations.

Dodia S., B. A., Mahesh P.A.

Cancer is considered to be a key cause of substantial fatality and morbidity in the world. A report from the International Agency for Research on Cancer (IARC) states that 27 million new cases of cancer are expected before 2030. 1 in 18 men and 1 in 46 women are estimated to develop lung cancer over a lifetime. This paper discusses an overview of lung cancer, along with publicly available benchmark data sets for research purposes. Recent research performed in medical image analysis of lung cancer using deep learning algorithms is compared using various technical aspects such as efficiency, advantages, and limitations. These discussed approaches provide insight into techniques that can be used to perform the detection and classification of lung cancer. Numerous techniques adapted in the acquisition of the images, extraction of relevant features, segmentation of region affected, selection of optimal features, and classification are also discussed. The paper is concluded by stating the clinical, technical challenges and prominent future directions.

Ojo M.O., Zahid A.

Controlled environment agriculture (CEA) is an unconventional production system that is resource efficient, uses less space, and produces higher yields. Deep learning (DL) has recently been introduced in CEA for different applications including crop monitoring, detecting biotic and abiotic stresses, irrigation, microclimate prediction, energy efficient controls, and crop growth prediction. However, no review study assess DL’s state of the art to solve diverse problems in CEA. To fill this gap, we systematically reviewed DL methods applied to CEA. The review framework was established by following a series of inclusion and exclusion criteria. After extensive screening, we reviewed a total of 72 studies to extract the useful information. The key contributions of this article are the following: an overview of DL applications in different CEA facilities, including greenhouse, plant factory, and vertical farm, is presented. We found that majority of the studies are focused on DL applications in greenhouses (82%), with the primary application as yield estimation (31%) and growth monitoring (21%). We also analyzed commonly used DL models, evaluation parameters, and optimizers in CEA production. From the analysis, we found that convolutional neural network (CNN) is the most widely used DL model (79%), Adaptive Moment Estimation (Adam) is the widely used optimizer (53%), and accuracy is the widely used evaluation parameter (21%). Interestingly, all studies focused on DL for the microclimate of CEA used RMSE as a model evaluation parameter. In the end, we also discussed the current challenges and future research directions in this domain.

Hussain M., Al-Aqrabi H., Munawar M., Hill R., Alsboui T.

Pallet racking is an essential element within warehouses, distribution centers, and manufacturing facilities. To guarantee its safe operation as well as stock protection and personnel safety, pallet racking requires continuous inspections and timely maintenance in the case of damage being discovered. Conventionally, a rack inspection is a manual quality inspection process completed by certified inspectors. The manual process results in operational down-time as well as inspection and certification costs and undiscovered damage due to human error. Inspired by the trend toward smart industrial operations, we present a computer vision-based autonomous rack inspection framework centered around YOLOv7 architecture. Additionally, we propose a domain variance modeling mechanism for addressing the issue of data scarcity through the generation of representative data samples. Our proposed framework achieved a mean average precision of 91.1%.

Zhang M., Yin L.

A solar cell defect detection method with an improved YOLO v5 algorithm is proposed for the characteristics of the complex solar cell image background, variable defect morphology, and large-scale differences. First, the deformable convolution is incorporated into the CSP module to achieve an adaptive learning scale and perceptual field size; then, the feature extraction capability of the model is enhanced by introducing the ECA-Net attention mechanism; finally, the model network structure is improved and one tiny defect prediction head is added to improve the accuracy of target detection at different scales. To further optimize and improve the YOLO v5 algorithm, this paper uses Mosaic and MixUp fusion data enhancement, K-means++ clustering anchor box algorithm, and CIOU loss function to enhance the model performance. The experimental results show that the improved YOLO v5 algorithm achieves 89.64% mAP for the model trained on the solar cell EL image dataset, which is 7.85% higher than the mAP of the original algorithm, and the speed reaches 36.24 FPS, which can complete the solar cell defect detection task more accurately while meeting the real-time requirements.

Hussain M., Al-Aqrabi H., Hill R.

This paper presents a framework for photovoltaic (PV) fault detection based on statistical, supervised, and unsupervised machine learning (ML) approaches. The research is motivated by a need to develop a cost-effective solution that detects the fault types within PV systems based on a real dataset with a minimum number of input features. We discover the appropriate conditions for method selection and establish how to minimize computational demand from different ML approaches. Subsequently, the PV dataset is labeled as a result of clustering and classification. The labelled dataset is then trained using various ML models before evaluating each based on accuracy, precision, and a confusion matrix. Notably, an accuracy ranging from 94% to 100% is achieved with datasets from two different PV systems. The model robustness is affirmed by performing the approach on an additional real-world dataset that exhibits noise and missing values.

Hussain M., Chen T., Hill R.

Pallet racking is a fundamental component within the manufacturing, storage, and distribution centers of companies around the World. It requires continuous inspection and maintenance to guarantee the protection of stock and the safety of personnel. At present, racking inspection is manually carried out by certified inspectors, leading to operational down-time, inspection costs and missed damage due to human error. As companies transition toward smart manufacturing, we present an autonomous racking inspection mechanism using a MobileNetV2-SSD architecture. We propose a solution that is affixed to the adjustable cage of a forklift truck, enabling adequate coverage of racking in the immediate vicinity. Our proposed approach leads to a classifier that is optimized for deployment onto edge devices, providing real-time alerts of damage to forklift drivers, with a mean average precision of 92.7%.

Li G., Jian X., Wen Z., AlSultan J.

Abstract This paper aims to eradicate the poor performance of the convolutional neural network (CNN) for intelligent analysis and detection in samples. Moreover, to avoid overfitting of the CNN model during the training process, an algorithm is proposed for the fusion of maximum pooled and weight decay. Firstly, the maximum pooled method for the pooling layer is explored after mask processing to reduce the number of irrelevant neurons. Secondly, when updating the neuron weight parameters, the weight decay is introduced to further cut down complexity in model training. The experimental comparison shows that the overfitting avoidance algorithm can reduce the detection error rate by more than 10% in image detection than other methods, and it has better generalisation.

Liu Z., Tan Y., He Q., Xiao Y.

Convolutional neural networks (CNNs) are good at extracting contexture features within certain receptive fields, while transformers can model the global long-range dependency features. By absorbing the advantage of transformer and the merit of CNN, Swin Transformer shows strong feature representation ability. Based on it, we propose a cross-modality fusion model, SwinNet , for RGB-D and RGB-T salient object detection. It is driven by Swin Transformer to extract the hierarchical features, boosted by attention mechanism to bridge the gap between two modalities, and guided by edge information to sharp the contour of salient object. To be specific, two-stream Swin Transformer encoder first extracts multi-modality features, and then spatial alignment and channel re-calibration module is presented to optimize intra-level cross-modality features. To clarify the fuzzy boundary, edge-guided decoder achieves inter-level cross-modality fusion under the guidance of edge features. The proposed model outperforms the state-of-the-art models on RGB-D and RGB-T datasets, showing that it provides more insight into the cross-modality complementarity task.

Ding X., Zhang X., Han J., Ding G.

We revisit large kernel design in modern convolutional neural networks (CNNs). Inspired by recent advances in vision transformers (ViTs), in this paper, we demonstrate that using a few large convolutional kernels instead of a stack of small kernels could be a more powerful paradigm. We suggested five guidelines, e.g., applying re-parameterized large depthwise convolutions, to design efficient high-performance large-kernel CNNs. Following the guidelines, we propose RepLKNet, a pure CNN architecture whose kernel size is as large as 31×31, in contrast to commonly used 3×3. RepLKNet greatly closes the performance gap between CNNs and ViTs, e.g., achieving comparable or superior results than Swin Transformer on ImageNet and a few typical downstream tasks, with lower latency. RepLKNet also shows nice scalability to big data and large models, obtaining 87.8% top-1 accuracy on ImageNet and 56.0% mIoU on ADE20K, which is very competitive among the state-of-the-arts with similar model sizes. Our study further reveals that, in contrast to small-kernel CNNs, large-kernel CNNs have much larger effective receptive fields and higher shape bias rather than texture bias. Code & models at https://github.com/megvii-research/RepLKNet.

Zhou H., Ahmed A., Liu T., Romeo M., Beh T., Pan Y., Kang H., Chen C.

Zhang W., Li X., Liu X., Lu S., Tang H.

Cao L., Wang Q., Luo Y., Hou Y., Zheng W., Qu H.

Khan A.T., Jensen S.M., Khan A.R.

Vezzali E., Bolelli F., Santi S., Grana C.

Xie S., Wu H., Mao W., Chu X., Meng Y., Yang X.

Sohrabi A., Ameli N., Mirimoghaddam M., Berlin-Broner Y., Lai H., Amin M.

Dental caries is a major global public health problem, especially among young children. Rapid decay progression often necessitates pulp treatment, making accurate pulp condition assessment crucial. Despite advances in pulp management techniques, diagnostic methods for assessing pulp involvement have not significantly improved. This study aimed to develop a machine learning (ML) model to diagnose pulp involvement using radiographs of carious primary molars. Clinical charts and bitewing radiographs of 900 children treated from 2018-2022 at the University of Alberta dental clinic were reviewed, yielding a sample of 482 teeth. images were preprocessed, standardized, and labeled based on clinical diagnoses. Data were split into training, validation, and test sets, with data augmentation applied to classify 2 categories of outcomes. The YOLOv8m-cls model architecture included convolutional and classification layers, and performance was evaluated using top-1 and top-5 accuracy metrics. The YOLOv8m-cls model achieved a top-1 accuracy of 78.7% for upper primary molars and 87.8% for lower primary molars. Validation datasets showed higher accuracy for lower primary teeth. Performance on new test images demonstrated precision, recall, accuracy, and F1-scores, highlighting the model’s effectiveness in diagnosing pulp involvement, with lower primary molars showing superior results. This study developed a promising CNN model for diagnosing pulp involvement in primary teeth using bitewing radiographs, showing promise for clinical application in pediatric dentistry. Future research should explore whole bitewing images, include clinical variables, and integrate heat maps to enhance the model. This tool could streamline clinical practice, improve informed consent, and assist in dental student training.

Ayhan B., Ayan E., Karadağ G., Bayraktar Y.

ABSTRACTObjectivesThe application of deep learning techniques for detecting caries in bitewing radiographs has gained significant attention in recent years. However, the comparative performance of various modern deep learning models and strategies to enhance their accuracy remains an area requiring further investigation.MethodsThis study explored the capabilities of 11 widely used YOLO (You Only Look Once) object detection models to automatically identify enamel and dentin caries from bitewing radiographs. To further optimize detection performance, the YOLOv9c model's backbone architecture was refined, reducing both model size and computational requirements. The enhanced model was assessed alongside six dentists, using the same test dataset for direct comparison.ResultsThe proposed YOLOv9c model achieved the highest performance among the evaluated models, with recall, precision, specificity, F1‐score, and Youden index values of 0.727, 0.651, 0.726, 0.687, and 0.453, respectively. Notably, the YOLOv9c model surpassed the performance of the dentists, as indicated by its recall and F1‐score values.ConclusionsThe proposed YOLOv9c model proved to be highly effective in detecting enamel and dentin caries, outperforming other models and even clinical evaluations by dentists in this study. Its high accuracy positions it as a valuable tool to augment dentists' diagnostic capabilities.Clinical SignificanceThe results emphasize the potential of the YOLOv9c model to assist dentists in clinical settings, offering accurate and efficient support for caries detection and contributing to improved patient outcomes.

Butorova A., Sergeev A.

Li S., Wang W., Lu Q.

Abstract Printed circuit boards (PCBs) with tiny defect detection face the problems of frequent omission and false detection, which seriously affect the reliability and safety of electronic products. To address these problems, a highly accurate and advanced PCB with a tiny defect detection model was proposed. This model incorporates four innovations. First, a small target detection head is augmented to capture more features for PCB with tiny defects. Second, the content-aware reassembly of features (CARAFE) operator is introduced to accumulate semantic information and local features. Third, a simple, parameter-free attention module (SimAM) is integrated into the C2f module to form the C2f-SimAM module, thereby strengthening the acquisition of channel and spatial information and enabling easier perception of tiny defects. Finally, space-to-depth with a non-strided convolution (SPD-Conv) module is used to dramatically reduce the loss of the feature map content. Therefore, this model is named CSS-YOLOv8. The results of this study confirm that the CSS-YOLOv8 model obtained a recall (R) of 95.5% with a mean average precision (mAP) of 97.9% on the PCB-DATASET dataset. After that, the CSS-YOLOv8 model had 7.2% and 5.7% improvements in R and mAP, respectively, compared to the original model. Accordingly, the CSS-YOLOv8 model significantly reinforces the accuracy of tiny defect detection in PCBs, and alleviates the omission and false detection of PCBs with tiny defects.

Cao G., Jia S., Wu Q., Xia C.

Mechanomyography (MMG) is a non-invasive technique for assessing muscle activity by measuring mechanical signals, offering high sensitivity and real-time monitoring capabilities, and it has many applications in rehabilitation training. Traditional MMG-based motion recognition relies on feature extraction and classifier training, which require segmenting continuous actions, leading to challenges in real-time performance and segmentation accuracy. Therefore, this paper proposes an innovative method for the real-time segmentation and classification of upper limb rehabilitation actions based on the You Only Look Once (YOLO) algorithm, integrating the Squeeze-and-Excitation (SE) attention mechanism to enhance the model’s performance. In this paper, the collected MMG signals were transformed into one-dimensional time-series images. After image processing, the training set and test set were divided for the training and testing of the YOLOv5s-SE model. The results demonstrated that the proposed model effectively segmented isolated and continuous MMG motions while simultaneously performing real-time motion category prediction and outputting results. In segmentation tasks, the base YOLOv5s model achieved 97.9% precision and 98.0% recall, while the improved YOLOv5s-SE model increased precision to 98.7% (+0.8%) and recall to 98.3% (+0.3%). Additionally, the model demonstrated exceptional accuracy in predicting motion categories, achieving an accuracy of 98.9%. This method realizes the automatic segmentation of time-domain motions, avoids the limitations of manual parameter adjustment in traditional methods, and simultaneously enhances the real-time performance of MMG motion recognition through image processing, providing an effective solution for motion analysis in wearable devices.

Mao M., Hong M.

Automated fabric defect detection is crucial for improving quality control, reducing manual labor, and optimizing efficiency in the textile industry. Traditional inspection methods rely heavily on human oversight, which makes them prone to subjectivity, inefficiency, and inconsistency in high-speed manufacturing environments. This review systematically examines the evolution of the You Only Look Once (YOLO) object detection framework from YOLO-v1 to YOLO-v11, emphasizing architectural advancements such as attention-based feature refinement and Transformer integration and their impact on fabric defect detection. Unlike prior studies focusing on specific YOLO variants, this work comprehensively compares the entire YOLO family, highlighting key innovations and their practical implications. We also discuss the challenges, including dataset limitations, domain generalization, and computational constraints, proposing future solutions such as synthetic data generation, federated learning, and edge AI deployment. By bridging the gap between academic advancements and industrial applications, this review is a practical guide for selecting and optimizing YOLO models for fabric inspection, paving the way for intelligent quality control systems.

Ren Y., Yu B., Xiu W., Ji H., Wang L.

Ta Q.T., Mac V.H., Huh J., Yim H.J., Lee G.

Liu Y., Huang Y., Bao Y.