Consistency　regularization　has　witnessed　tremendous　success　in　the　area　of　semi-supervised　deep　learning　for　image　classification,　which　leverages　data　augmentation　on　unlabeled　examples　to　encourage　the　model　outputting　the　invariant　predicted　class　distribution　as　before　augmented.　These　methods　have　been　made　considerable　progress　in　this　area,　but　most　of　them　are　at　the　cost　of　utilizing　more　complex　models.　In　this　work,　we　propose　a　simple　and　efficient　method　FMixAugment,　which　combines　the　proposed　MixAugment　with　Fourier　space-based　data　masking　and　applies　it　on　unlabeled　examples　to　generate　a　strongly-augmented　version.　Our　approach　first　generates　a　hard　pseudo-label　by　employing　a　weakly-augmented　version　and　minimizes　the　cross-entropy　between　it　and　the　strongly-augmented　version.　Furthermore,　to　improve　the　robustness　and　uncertainty　measurement　of　the　model,　we　also　enforce　consistency　constraints　between　the　mixed　augmented　version　and　the　weakly-augmented　version.　Ultimately,　we　introduce　a　dynamic　growth　of　the　confidence　threshold　for　pseudo-labels.　Extensive　experiments　are　tested　on　CIFAR-10/100,　SVHN,　and　STL-10　datasets,　which　indicate　that　our　method　outperforms　the　previous　state-of-the-art　methods.　Specifically,　with　40　labeled　examples　on　CIFAR-10,　we　achieve　90.21%　accuracy,　and　exceed　95%　accuracy　with　1000　labeled　examples　on　STL-10.　©　2021,　Springer　Nature　Switzerland　AG.