Solar　cells　are　the　fundamental　core　energy　harvesting　components　in　photovoltaic　(PV)　power　generation　stations.　In　view　of　the　capability　of　detecting　the　invisible　defects,　the　electroluminescence　(EL)　imaging　is　broadly　used　in　the　production　lines　of　solar　cells,　based　on　which　the　deep　learning　technique　is　introduced　to　implement　automatic　defect　detection　and　classification.　However,　the　current　deep　learning　models　feature　high　complexity　and　require　much　computation　resources,　which　are　difficult　to　deploy　in　edge　devices　for　real　time　applications.　To　tackle　this　issue,　we　proposed　a　novel　lightweight　and　high-precision　deep　learning　model　named　Cross　Stage　Partial　Photovoltaic-You　Only　Look　Once　(CSPV-YOLO)　based　on　the　deep　learning　framework　You　Only　Look　Once　v5　(YOLOv5)　to　enable　the　real-time　solar　cell　defect　detection.　Firstly,　a　new　module　Cross　Stage　Partial　C5　(CSPC5)　is　proposed　to　replace　the　initial　C3　module　in　the　YOLOv5　network　to　enhance　the　network　accuracy　in　recognizing　different　types　of　defects.　Secondly,　a　novel　Spatial　Pyramid　Pooling　with　Cross　Stage　Partial　(SPPFCSP)　module　is　designed　to　replace　the　original　Spatial　Pyramid　Pooling　Fast　(SPPF),　which　boosts　the　network　feature　extraction　capabilities　from　defect　targets　at　multiple　scales　and　facilitates　a　more　efficient　integration　of　multiscale　features.　Finally,　the　original　loss　function　of　YOLOv5　is　replaced　by　the　Scylla　intersection　over　union　(SIoU)　function　to　optimize　the　training　model.　The　proposed　models　have　been　validated　and　intensively　compared　with　many　other　state-of-the-art　models　on　two　public　datasets.　Firstly,　results　of　experiments　on　the　public　Pascal　Visual　Object　Classes　(PASCAL　VOC)　2007　datasets　demonstrate　that　the　proposed　SPPFCSP　block　is　obviously　superior　to　other　Spatial　Pyramid　Pooling　blocks　for　the　most　state-of-the-art　YOLO　detectors,　which　can　significantly　improve　the　detection　accuracy.　The　comparison　results　of　experiments　on　the　public　Photovoltaic　Electroluminescence　Anomaly　Detection　Dataset　(PVEL-AD)　that　includes　12-class　defects　obviously　indicate　that　the　proposed　CSPV-YOLO　model　is　better　than　many　state-of-the-art　models　and　achieves　91.5　%　average　precision　(AP)　and　frames　per　second　(FPS)　of　177.8　on　with　only　2.2　million　(M)　parameters.　Hence,　it　is　suitable　for　the　deployment　on　edge　devices　for　real-time　applications.　©　2025　Elsevier　Ltd