Curved　mirrors　are　integral　components　for　achieving　high-quality　projections　in　lens　modules,　where　design,　manufacturing,　and　installation　errors　can　significantly　affect　projection　quality.　Therefore,　it　is　essential　to　precisely　adjust　the　mirror＇s　installation　position　based　on　the　distortion　characteristics　of　the　projected　image　before　finalizing　the　assembly　of　the　lens,　ensuring　optimal　projection　quality.　This　study　addresses　the　challenges　associated　with　the　precise　calibration　of　mirrors　during　the　assembly　process　of　large-format　ultra-short　throw　laser　projectors　by　developing　a　rapid　automatic　mirror　adjustment　system.　Considering　the　spatial　constraints　and　limited　adjustability　of　mirrors　within　the　lens　module,　along　with　the　dark　calibration　environment,　we　designed　a　mechanical　adjustment　system　comprising　a　specialized　positioning　device　for　the　lens　module,　a　three-dimensional　motion　platform,　and　a　two-degree-of-freedom　end　effector.　To　ensure　accurate　capture　of　the　adjusting　nut　of　the　mirror　by　the　end　effector,　we　adopted　an　improved　Yolov5-swin-ECA　algorithm,　which　employs　the　Swin　Transformer　as　the　backbone　of　the　Yolov5　architecture　and　integrates　an　ECA　attention　mechanism　within　the　feature　extraction　network.　Assisted　by　a　depth　camera,　this　approach　enables　rapid　and　precise　positioning.　To　tackle　the　challenges　posed　by　the　high　precision　requirements　and　extensive　image　capture　area　in　ultra-short　throw　projection　screens,　we　implemented　a　four-camera　segmented　acquisition　scheme　that　facilitates　the　collection　of　refined　data　on　the　projected　images.　Using　the　initial　distortion　image　sample　set　provided　by　our　industrial　partner,　the　automatic　adjustment　system　can　reversely　control　standard　images　to　simulate　distorted　ones,　obtaining　the　relationship　between　the　correction　amounts　of　the　end　effector　and　the　image　distortion—this　process　establishes　a　distortion　correction　dataset　for　training　the　automatic　adjustment　system.　For　addressing　random　image　distortion　automatic　adjustment,　an　improved　Transformer-KAN　model　was　introduced,　which　learns　from　the　collected　dataset　to　determine　the　necessary　rotational　adjustments　required　for　the　end　effector　to　transform　arbitrarily　distorted　images　into　standard　images.　Through　iterative　deep　learning　training,　the　post-processing　system　develops　self-learning　capabilities,　rapidly　executing　automatic　mirror　adjustments　and　significantly　enhancing　the　efficiency　and　quality　of　the　curvature　mirror　calibration.　©　2025