A　fast　and　effective　approach　of　automatic　text　localization　in　images　with　complex　background　is　presented.　Text　areas　can　be　segmented　with　an　adaptive　threshold　using　a　new　feature　extracted　from　DCT-compressed　domain,　and　then　a　coarse-to-fine　projecting　and　merging　algorithm　is　proposed　to　locate　the　final　text　area.　Experiments　based　on　JPEG　image　testing　sets　show　its　effective　and　robustness　for　contrast,　language,　and　background　complexity.