Laboratory　safety　education　is　fundamental　to　the　smooth　conduct　of　scientific　research.　Traditional　educational　models　are　often　limited　in　interactivity　and　real-time　feedback　and　they　struggle　to　satisfy　the　increasing　needs　of　laboratory　safety　management.　This　study　aims　to　enhance　the　interactivity　and　efficacy　of　laboratory　safety　education　by　adopting　machine　learning-enhanced　image　processing　technologies.　To　cope　with　the　noise　issue　in　hazardous　behavior　data　within　laboratories,　wavelet　threshold　denoising　is　applied　to　significantly　improve　data　usability.　To　deal　with　the　imbalance　in　data　samples　of　laboratory　hazard　behaviors,　an　adaptive　boundary　data　augmentation　algorithm　is　introduced　to　balance　the　dataset　and　strengthen　the　model＇s　generalization　capability.　A　breakthrough　in　extracting　spatio-temporal　features　of　hazardous　behaviors　in　laboratories　is　achieved　through　an　improved　Spatio-Temporal　Graph　Convolutional　Network　(ST-GCN)　model,　enabling　effective　recognition　and　classification　of　hazardous　behaviors.　The　outcome　attained　in　this　study　has　important　implications　for　enhancing　the　interactivity　and　practicality　of　laboratory　safety　education　and　it　can　expand　new　research　directions　for　the　application　of　machine　learning　in　the　field　of　image　processing.　©2023　IIETA.　This　article　is　published　by　IIETA　and　is　licensed　under　the　CC　BY　4.0　license.