A Photovoltaic Cell Defect Detection Method Based on Improved RT-DETR

Abstract

In photovoltaic cell manufacturing, various surface defects including cracks, black cores, and finger interruptions may be introduced during production and operation, leading to reduced power generation efficiency, accelerated degradation, and potential module failure. Although deep learning-based methods have advanced automated inspection, existing defect detection methods still suffer from insufficient representation of weaktexture defects and limited feature discrimination capability under complex backgrounds. To address these issues, we propose a lightweight detection method based on an improved RT-DETR (Real-Time Detection Transformer) model. Specifically, a reparameterized dynamic partial convolution backbone (RDPCNet) is designed for efficient lightweight feature extraction. It leverages channel splitting for computational reduction and contentadaptive gating calibration for fine-grained defect feature extraction. Meanwhile, an adaptive context enhancement module (ACG-Block) is constructed, which enhances multi-scale feature fusion through a configurable pooling pyramid and coordinate attention for position-sensitive spatial refinement. Extensive experiments conducted on the PVEL-AD photovoltaic cell defect dataset demonstrate that the proposed model improves mAP50 by 2.5 percentage points while reducing parameters by 23.4% and GFLOPs by 15.4% compared to the baseline RT-DETR. These results indicate the distinct advantages of the proposed approach in detecting surface defects in photovoltaic cells.