The　time　and　effort　required　to　manually　design　deep　neural　architectures　is　extremely　high,　which　has　led　to　the　development　of　neural　architecture　search　technology　as　an　automatic　architecture　design　method.　However,　the　neural　architecture　search　convergence　process　is　slow　and　expensive,　and　the　process　requires　training　a　large　number　of　candidate　architectures　to　get　the　final　result.　If　the　final　accuracy　of　an　architecture　can　be　predicted　from　its　initial　state,　this　problem　can　be　greatly　alleviated.　Therefore,　this　paper　proposes　a　low-cost　architecture　performance　evaluation　strategy　based　on　pixel　difference　degree　contrast　measurement,　which　takes　1)　the　difference　matrix　value　between　the　feature　map　generated　in　the　untrained　architecture　and　the　original　image,　and　2)　the　predicted　accuracy　of　the　neural　network　as　evaluation　indices.　A　new　multi-index　weight　comprehensive　measurement　strategy　was　introduced　to　comprehensively　score　the　multi-index,　the　real　architecture　performance　can　be　approximately　represented　by　score,　which　greatly　reduces　the　cost　of　architecture　evaluation.　The　experimental　show　that　the　proposed　scoring　strategy　is　highly　correlated　with　real　architecture　accuracy.　In　the　practical　engineering　application　research,　this　strategy　can　search　a　high-performance　architecture　with　an　accuracy　of　96.2%　within　343.3　s,　which　proves　that　the　proposed　strategy　can　significantly　improve　the　search　efficiency　in　practical　applications,　reduce　the　subjectivity　of　artificial　architecture　design,　and　promote　the　application　of　practical　time-consuming　projects.　©　2024　Elsevier　B.V.