Attention classification-and-segmentation network for micro-crack anomaly detection of photovoltaic module cells

Abstract

Micro-crack anomaly detection is a crucial part of the quality inspection of photovoltaic (PV) module cells. However, due to the complex background and the lack of sufficient anomaly samples, it is a challenging task to identify and locate micro-crack accurately. This paper presents a novel method for detecting micro-crack anomaly in PV module cells by designing an attention classification-and-segmentation network. Specifically, the proposed network consists of a classification network and a segmentation network. In the classification network, the real-time micro-crack anomaly discrimination task in the detection process is performed at first. The classification network introduces transfer learning and deep supervision mechanism to effectively extract and fuse multi-scale features to accurately predict the anomaly probability score. In the segmentation network, the pixel-level micro-crack detection is conducted on the sample determined as defects. The M-shaped structure is used to better extract and fuse shallow-level and deep-level features in the segmentation network, effectively solving the “All Black” issue in the training process. And the attention mechanism module is inserted into the M-shaped structure to more effectively extract micro-crack anomaly features and suppress background noise, thereby significantly improving the accuracy of segmentation. By the design of two-stage network architecture and the utilization of the attention module, the proposed network presents a strong capability for learning from a small set of labeled and annotated samples. Comprehensive experiments are conducted on the real PV electroluminescence (EL) images dataset. Experimental results show that, compared with InceptionV4 classification network and U-net segmentation network, the proposed network has superior performance with ACC of 100.0% and DICE of 0.541 in the micro-crack anomaly classification and segmentation task. Moreover, the experiments show the attention module inserted in the proposed network plays a significant role in improving the classification and segmentation accuracy.