針對(duì)圖像指代分割的訓(xùn)練后量化策略

打開(kāi)文本圖片集

關(guān)鍵詞：圖像指代分割；訓(xùn)練后量化；跨模態(tài)融合；深度學(xué)習(xí)

中圖分類(lèi)號(hào)：TP391.41;TP183 文獻(xiàn)標(biāo)志碼：A 文章編號(hào)：1001-3695（2025）07-014-2025-07

doi：10.19734/j. issn.1001-3695.2024.10.0437

Abstract：RISaims tosegmentobjectsdescribedbysentencesinanimagebyintegratingvisualandlinguisticinformation. This technique has strong appication prospects ininteractiveimage editingandlanguage-guided human-machine interaction. However，existing solutions tendtoexplore high-performance models，neglecting practicalapplicationsonedgedeviceswith limited esources.ThepaperproposedaneficientPQframework toaddressthischallenge.Specifically，theanalysisdentifiedtherotcauseofperformancecollpsecausedbyusingtheround-to-nearest（RTN）quantization method.Thentheframework proposedatwo-regionbalancedquantizationstrategytosolvethenon-normaldistributionofactivationvaluesaftersoftmax and GELUoperations inthevisual encoder，andintroducedareordered groupingquantization strategytotacklethequantizationproblemscausedbyoutliersinthelinearlayersactivationof the textencoder.Extensiveexperimentsexploringdierent quantization bitwidthsonthreebenchmark datasetsdemonstratetheclearadvantages ofthe proposed methodover existing PTQ methods.AsthefirstquantizationschemespecificallfortheRIStask，itconfirmsthefeasibilityofeficientlydeployingthe RIS model to edge devices using the PTQ method.

Key words： referring image segmentation（RIS）； post-training quantization（PTQ）;cross-model fusion; deep learning

0引言

深度學(xué)習(xí)極大程度提高了視覺(jué)算法在許多圖像分割任務(wù)上的性能，如語(yǔ)義分割[1]實(shí)例分割[2]等。（剩余18265字）