Quantitative　detection　of　defects　in　structures　is　always　a　hot　research　topic　in　the　field　of　guided　wave　inverse　scattering.　Research　studies　on　how　to　effectively　extract　the　defect-related　information　encompassed　in　the　multi-frequency　and　multi-modes　scattered　wave　signals　for　reconstructions　of　defects　have　been　paid　attention　in　recent　decades.　In　this　paper,　a　novel　deep　learning-based　quantitative　guided　wave　inverse　scattering　technique　has　been　proposed　to　intelligently　realize　the　end-to-end　mapping　of　the　multi-frequency,　multi-modes　scattered　signals　to　defect　profiles　with　high　levels　of　accuracy　and　efficiency.　Based　on　the　manifold　distribution　principle,　the　data　patterns　of　scattered　SH-wave　signals　have　been　investigated,　owing　to　leveraging　the　capability　of　the　intelligent　encoder-projection-decoder　neural　network.　Following　that,　the　manifold-learning-oriented　network　has　been　trained　using　the　data　generated　by　the　modified　boundary　element　method.　Several　numerical　examples　have　been　examined　to　demonstrate　the　correctness　and　efficiency　of　the　proposed　reconstruction　approach.　It　has　been　concluded　that　this　novel　data-driven　technique　intelligently　enables　the　high-quality　solution　to　inverse　scattering　problems　and　provides　valuable　insight　into　the　development　of　practical　approaches　to　quantitative　detection　using　multi-frequency　and　multi-modal　acoustic　data　from　scattered　ultrasonic　guided　waves.