In　the　evolution　of　non-contact　human-computer　interactions,　the　integration　of　low-resolution　infrared　thermal　pile　sensors　has　gained　prominence,　particularly　in　applications　like　smart　homes　and　human-computer　collaboration.　However,　challenges　arise　from　uncontrollable　environmental　temperatures,　causing　issues　such　as　blurry　or　missing　details　in　finger　contours　during　infrared　thermal　pile　imaging.　This　paper　addresses　these　challenges　by　employing　the　DiffBIR　diffusion　model　for　gesture　reconstruction　and　integrating　the　CBAM　module　into　a　lightweight　gesture　recognition　network　named　Infrared　Image　Reconstruction　Gesture　Network　(IR-GNet).　The　proposed　approach　enhances　attention　weights　for　hand　regions　and　effectively　eliminates　interference　from　other　heat　sources,　ensuring　optimal　performance　in　low-resolution　scenarios.　Our　model　achieved　state-of-the-art　accuracy　on　both　static　and　dynamic　gesture　datasets,　and　on　the　Worker-Robot　Collaboration　(WRC)　gesture　dataset.　Experiments　involving　thermal　source　interference　also　demonstrated　that　the　model　outperforms　other　lightweight　models　in　terms　of　recognition　accuracy　while　meeting　real-time　requirements.　Moreover,　this　study　introduces　a　joint　training　approach　to　mitigate　the　relatively　high　time　cost　associated　with　image　reconstruction.　By　combining　the　training　set　of　original　images　with　their　corresponding　reconstructed　images,　and　using　only　original,　unreconstructed　images　for　testing,　significant　enhancements　in　recognition　accuracy　were　observed　under　identical　testing　conditions.　These　results　affirm　the　feasibility　of　deploying　the　proposed　network　in　real-time　scenarios　to　achieve　a　robust　gesture　recognition　system.