Aspect-level　multimodal　sentiment　analysis　aims　to　ascertain　the　sentiment　polarity　of　a　given　aspect　from　a　text　review　and　its　accompanying　image.　Despite　substantial　progress　made　by　existing　research,　aspect-level　multimodal　sentiment　analysis　still　faces　several　challenges:　(1)　Inconsistency　in　feature　granularity　between　the　text　and　image　modalities　poses　difficulties　in　capturing　corresponding　visual　representations　of　aspect　words.　This　inconsistency　may　introduce　irrelevant　or　redundant　information,　thereby　causing　noise　and　interference　in　sentiment　analysis.　(2)　Traditional　aspect-level　sentiment　analysis　predominantly　relies　on　the　fusion　of　semantic　and　syntactic　information　to　determine　the　sentiment　polarity　of　a　given　aspect.　However,　introducing　image　modality　necessitates　addressing　the　semantic　gap　in　jointly　understanding　sentiment　features　in　different　modalities.　To　address　these　challenges,　a　multi-granularity　visual-textual　feature　fusion　model　(MG-VTFM)　is　proposed　to　enable　deep　sentiment　interactions　among　semantic,　syntactic,　and　image　information.　First,　the　model　introduces　a　multi-granularity　hierarchical　graph　attention　network　that　controls　the　granularity　of　semantic　units　interacting　with　images　through　constituent　tree.　This　network　extracts　image　sentiment　information　relevant　to　the　specific　granularity,　reduces　noise　from　images　and　ensures　sentiment　relevance　in　single-granularity　cross-modal　interactions.　Building　upon　this,　a　multilayered　graph　attention　module　is　employed　to　accomplish　multi-granularity　sentiment　fusion,　ranging　from　fine　to　coarse.　Furthermore,　a　progressive　multimodal　attention　fusion　mechanism　is　introduced　to　maximize　the　extraction　of　abstract　sentiment　information　from　images.　Lastly,　a　mapping　mechanism　is　proposed　to　align　cross-modal　information　based　on　aspect　words,　unifying　semantic　spaces　across　different　modalities.　Our　model　demonstrates　excellent　overall　performance　on　two　datasets.