With　the　growing　concern　for　power-hungry　on　mobile　devices,　many　power　constrained　contrast　enhancement　algorithms　have　been　developed　in　the　mobile　devices　embedded　with　emissive　displays,　such　as　organic　light-emitting　diodes.　However,　conventional　power　constrained　contrast　enhancement　algorithms　inevitably　degrade　the　visual　aesthetics　of　images　as　a　trade-off　to　gain　the　power-saving　for　mobile　devices.　This　paper　proposes　a　trainable　power-constrained　contrast　enhancement　algorithm　based　on　a　saliency-guided　deep　framework　for　suppressing　the　power　consumption　of　an　image　while　preserving　its　perceptual　quality.　Our　algorithm　relies　on　the　fact　that　imaging　features　of　a　displayed　image　is　salient　to　human　visual　perception.　Hence,　we　decompose　the　input　image　into　the　imaging　features　and　textual　features　with　a　deep　convolutional　neural　networks,　and　degrade　those　textual　features　to　achieve　the　suppression　of　power　consumption.　Experimental　results　demonstrate　that　our　algorithm　is　able　to　maintain　visual　aesthetics　of　images　while　reducing　the　power　consumption　effectively,　outperforming　conventional　power-constrained　contrast　enhancement　algorithms.　©　2018　IEEE.