In　recent　research,　the　defect　detection　algorithm　based　on　the　fully-supervised　object　detection　model　has　become　one　of　the　research　hotspots　and　has　achieved　good　results.　However,　fully-supervised　object　detection　models　require　image-level　and　localization-level　labels.　Obtaining　these　labels　requires　a　great　deal　of　manpower.　Therefore,　this　paper　proposes　a　dual　path　defect　detection　network　(DPNET)　based　on　weakly　supervised　object　detection　model,　which　aims　to　identify　the　classification　label　and　carry　on　localization　for　defects　merely　by　using　image-level　labels.　Firstly,　the　paper　employs　the　deep　convolutional　residual　network　ResNet-50　as　a　feature　classification　network　for　defect　classification.　Secondly,　we　designed　a　localization　network　based　on　the　global　average-max　pooling　class　activation　map　(GAM-CAM)　and　the　Full　Convolutional　Channel　Attention　(FCCA)　for　defect　localization,　which　can　improve　the　defect　localization　accuracy.　Experimental　results　on　the　DAGM　dataset　confirm　that　the　proposed　detection　model　is　able　to　efficiently　detect　defects.　©　2021,　Springer　Nature　Switzerland　AG.