The　catastrophic　forgetting　problem　is　one　of　the　hotspots　in　the　field　of　deep　learning.　At　present,　there　is　no　doubt　that　storing　samples　of　previous　tasks　in　fixed-size　memory　is　the　best　way　to　solve　this　problem.　However,　the　number　of　samples　stored　in　fixed-size　memory　is　limited.　With　the　increase　in　tasks,　the　number　of　samples　stored　in　memory　for　a　single　task　will　decrease　sharply.　It　is　also　difficult　to　balance　the　memory　capacity　and　number　of　samples.　To　solve　this　problem,　some　methods　use　a　fixed　size　of　memory　to　store　dimensionality　reduction　images.　However,　this　will　create　new　problems.　1)　The　quality　of　dimensionality　reduction　images　is　poor,　and　they　are　significantly　different　from　original　images.　2)　How　to　choose　the　dimensionality　reduction　method　of　images.　To　address　these　problem,　we　put　forward　a　new　method.　Firstly,　we　employ　a　simple　and　reliable　scheme　to　solve　the　domain　difference　between　dimensionality　reduction　images　and　original　images.　And　we　theoretically　analyzed　which　image　dimensionality　reduction　method　is　better.　Secondly,　to　increase　the　generalization　ability　of　our　method　and　further　mitigate　the　catastrophic　forgetting　phenomenon,　we　utilize　a　self-supervised　image　augmentation　method　and　the　output　features　similarity　loss.　Thirdly,　we　make　use　of　the　neural　kernel　mapping　support　vector　machine　theory　to　improve　the　interpretability　of　our　method.　Experimental　results　demonstrated　that　the　top-1　average　accuracy　of　our　method　is　much　higher　than　other　methods　when　using　the　same　size　of　memory.