Shenshen Zhao

Developing reliable, generalized, and accurate defect detection technology for photovoltaic (PV) manufacturers is particularly critical with the demand for production line expansion. Existing technologies perform well when handling independently and identically distributed (IID) data. However, their performance significantly reduces when they encounter domain shift problems, including style and instance bias, prompted by production line expansion. In this paper, we propose a novel Dual Bias Adjustment Network (DBAN) to enhance the generalization and reliability of PV defect detection. Specifically, we construct a Global Style-Generalized Contrastive Learning (GSCL), which uses nonlinear transformation functions and contrastive learning strategies to enhance model adaptability for different style changes and global discriminative ability, effectively overcoming the style bias problem. We design a Test-Time Prototype Adjustment (TTPA) that employs graph methods and prototype learning to adjust feature representations accurately. TTPA enhances prediction reliability during testing via dynamic prototype repositories and memory mechanisms, effectively addressing instance deviations. We conduct comprehensive experiments proving that DBAN achieves optimal performance, surpassing other advanced algorithms. Moreover, GSCL and TTPA add little inference time to the model, making them suitable for practical industrial applications. Finally, we conduct extensive experiments on the public domain-shifted PV dataset ELES, where our model achieves state-of-the-art performance in Single-domain generalized object detection.

It is significant to accurately identify the risk target of external force damage to ensure the safe operation of transmission lines. The recent object detection networks have provided ideas for determining the external force destruction risk object, and they have all fused deep and shallow features to a certain extent in the network. However, the fusion of features with different depths without processing will cause the current model to lose risk targets in weak areas. In this paper, we propose a new Cross-Phase Fusion (CPF) detector to detect small and weak risk targets effectively, which embeds a Cross-Fusion Module of Graph Reasoning Based on Stripe Pooling (CGRS) in specific layers. Specifically, CGRS adaptively captures random spatial dependence and long-range channel dependence with directional information to enhance the extraction ability of spatial features. On the other hand, CGRS utilizes an interactive projection mechanism to project spatial and semantic information into the graph space effectively. The reasoning between graph nodes realizes the dynamic correlation of target-related features and improves the feature representation of small targets. We conducted comprehensive experiments to prove that the value of mAP0.5 reached 80.1% in training external risk datasets for the CPF detector, which is superior to the current advanced methods and verifies its effectiveness for detecting external force damage risk. We believe this method can have practical application value in maintaining the safety of transmission lines.

Aeroengine blades are critical components in engines, and any minor surface damage may lead to severe consequences. Therefore, high-precision damage segmentation techniques are crucial for accurately identifying blade damage. However, the high-precision localization of blade damages faces challenges due to the loss of detailed spatial features and the confusion of semantic features. In this paper, we propose a Spatial Contrast and Semantic Difference Perception Network (SSPN) for aeroengine blade damage segmentation, which effectively enhances the discriminative ability of spatial and semantic feature representations. Specifically, we propose a Class-aware Spatial Contrast (CSC) module to first generate segmentation masks and highlight categories. Then, CSC utilizes a class-aware contrastive learning method to enhance the distinguishability of spatial features. To establish clear semantic boundaries, we design a Semantic Difference Perception (SDP) module that first extracts semantic difference information between feature maps. Then, SDP uses a difference perception loss function to focus on low-quality areas and capture subtle semantic variances, which can produce a more balanced and contrasting semantic feature representation. We conduct comprehensive experiments to demonstrate the superiority of SSPN that achieves 93.42% mIoU on our blade dataset, surpassing existing state-of-the-art methods. Furthermore, we build an Aeroengine Blade Damage Analysis System to formulate more rational and economical maintenance strategies that can facilitate the development of the aviation engineering field.

The photovoltaic industry is a strategic industry with international competitive advantages and is developing toward a larger scale, higher efficiency, and higher quality. However, current researchers have not built a pixel-level defect inspection system for large-scale photovoltaic production processes. This article proposes an intelligent defect segmentation system combining the Internet of Things (IoT), artificial intelligence (AI), and edge computing for quality inspection of large-scale photovoltaic production lines. The intelligent factory based on this system is highly intelligent and deeply integrated, which can significantly reduce labor costs and improve factory productivity and product quality. The system’s core uses edge computing to segment cells in real-time by a lightweight defect segmentation model. Specifically, this article proposes a lightweight yet effective architecture named low-cost defect segmentation network (LDSN). An efficient split (ES) block is designed to support more channels and improve model accuracy without adding much computational complexity. Moreover, the ES block can express multiscale features in a finer granularity and enhance the information interaction between grouping features. In the decoding structure, a dual focus attention (DFA) that efficiently captures long-range spatial and channel information is proposed. Comprehensive experiments have been performed on a low-end PC with an NVIDIA GeForce RTX3060 GPU and an Intel Core i5-10600KF. LDSN-T-Lite achieves 84FPS and the F-measure optimal image scale (OIS) of 0.827, which only has 166K parameters and 395.6M memory usage on our PSCDE data set. A bigger version of LDSN-B achieves the F-measure OIS of 0.872, significantly outperforming current methods.

Most of the existing photovoltaic (PV) defect detection methods are based on the assumption that the training and testing samples satisfy the independent identically distributed. However, in real PV scenarios, the endogenous shift problem exists widely, including background style shift and defect instance shift, which seriously affects the performance of detectors. In this article, we propose a novel network called shift suppression network (SSN) for the endogenous shift of PV defect detection, which consists of two core components, background style suppression (BSS) module and cross-layer graph reasoning (CGR) module. Specifically, BSS uses channel statistics matching alignment to adaptively suppress background style shifts without reliance on unknown domain data. CGR learns semantic dependencies between multiscale channel feature maps through cross-layer interaction, which improves the discriminative ability and localization ability in the face of defect instance shifts. To advance the study of endogenous shift, we provide the first electroluminescence (EL) endogenous shift dataset for PV modules, which creates by three groups of EL images collected at different times, totaling 16 323. The comprehensive evaluation results show that the SSN outperforms the state-of-the-art methods. Furthermore, we design an intelligent defect detection system for PV modules based on SSN. The system has higher detection accuracy and an intuitive visual interface, which can meet the actual detection function and requirements of the production line.

High-performance defect segmentation techniques are essential for the high-quality manufacturing of polycrystalline solar cells. Edge detection is an effective technique to accurately locate the edge of defects. However, the existing methods ignore global channel information and the representation gap between multiscale features, inhibiting the ability of the network to aggregate discriminative features. In this article, we propose a novel rich edge features refinement detection network consisting of an encoder–decoder structure that captures rich discriminative edge feature representations by interactively exploring global spatial and channel context. To refine the dense global contextual information in the decoding layer, we propose a bidirectional strip refinement attention (BSRA) adaptively capturing long-range spatial dependency with direction information and long-range channel dependency. BSRA is a lightweight and general module that can be easily inserted into the existing edge detection networks with a negligible computational burden. In addition, we release a polycrystalline solar cell defect edge (PSCDE) dataset that is the first high-quality segmentation database to advance the development of high-quality polycrystalline solar cell manufacturing. Our method achieves an ODS F-measure of 0.854 on the PSCDE dataset with strict evaluation criteria (maxDist = 0.0015), outperforming existing state-of-the-art methods. To further verify the generalization ability of BSRA, we apply BSRA to other edge detection networks, and experiments show that the module further improves the accuracy of these methods on PSCDE.

High-quality and fast crack segmentation in solar cell electroluminescence images with heterogeneously textured background disturbance is challenging. We propose an end-to-end Efficient and Refined Deep Convolutional Features Network (ERDCF-Net) for precise and efficient crack segmentation. Firstly, we design a lightweight, efficient deep convolutional features (EDCF) encoder by adopting multilevel deep convolutional features to generate a discriminative feature representation of cracks. Moreover, some complicated background disturbance is restrained by reusing rich crack and background features. Secondly, we present a refined side output (RS) decoder by automatically refining side outputs from the EDCF encoder. The experiments demonstrate that the proposed network excellently performs on the solar cell cracks (MIoU of 92.82%, F -measure of 93.58%, and 89 FPS) and open crack datasets (MIoU of 85.6%, F -measure of 84.1%, and 90 FPS).