Development of web application for try on clothes
Jazyk: | ruština |
---|---|
Rok vydání: | 2022 |
Předmět: | |
DOI: | 10.18720/spbpu/3/2022/vr/vr22-2791 |
Popis: | ÐиÑÑÑалÑÐ½Ð°Ñ Ð¿ÑимеÑка â пеÑÑпекÑивное и бÑÑÑÑоÑаÑÑÑÑее напÑавление глÑбокого обÑÑениÑ, коÑоÑое напÑавлено на пÑеобÑазование внеÑнего вида Ñеловека пÑÑÑм Ð´Ð¾Ð±Ð°Ð²Ð»ÐµÐ½Ð¸Ñ Ð½Ð¾Ð²ÑÑ Ð¿Ñизнаков (макиÑжа, паÑиков), а Ñакже Ð¸Ð·Ð¼ÐµÐ½ÐµÐ½Ð¸Ñ ÑÑÑеÑÑвÑÑÑÐ¸Ñ (одеждÑ). Рданной ÑабоÑе ÑаÑÑмоÑÑÐµÐ½Ñ Ð½Ð°Ð¸Ð±Ð¾Ð»ÐµÐµ попÑлÑÑнÑе и знаÑимÑе меÑÐ¾Ð´Ñ Ð²Ð¸ÑÑÑалÑной пÑимеÑки одеждÑ. РкаÑеÑÑве оÑновного алгоÑиÑма бÑла вÑбÑана нейÑÐ¾Ð½Ð½Ð°Ñ ÑеÑÑ PF-AFN, демонÑÑÑиÑÑÑÑÐ°Ñ Ð½Ð°Ð¸Ð»ÑÑÑие ÑезÑлÑÑаÑÑ Ð¿Ñи наименÑÑÐ¸Ñ ÑÑебованиÑÑ Ðº Ð²Ñ Ð¾Ð´Ð½Ñм даннÑм. РпÑоÑеÑÑе ÑабоÑÑ Ð½Ð°Ð´ вÑбÑаннÑм алгоÑиÑмом бÑло найдено множеÑÑво огÑаниÑений ÑÑÑеÑÑвÑÑÑей ÑеализаÑии, в ÑвÑзи Ñ Ñем она бÑла модиÑиÑиÑована Ñ ÑÑÑÑом ÑовÑеменнÑÑ Ð¼ÐµÑодов опÑимизаÑии, ÑÐ°ÐºÐ¸Ñ ÐºÐ°Ðº обÑÑение Ñо ÑмеÑанной ÑоÑноÑÑÑÑ Ð¸ ÑаÑпÑеделÑнное ÑаÑпаÑаллеливание даннÑÑ . ÐÑи ÑÑом опÑимизаÑÐ¸Ñ Ð½Ðµ повлиÑла на каÑеÑÑво иÑогового ÑезÑлÑÑаÑа, как показали вÑÑиÑленнÑе меÑÑики. ÐÐ»Ñ Ð´ÐµÐ¼Ð¾Ð½ÑÑÑаÑии пÑименимоÑÑи меÑода в пÑомÑÑленнÑÑ ÑеÑениÑÑ Ð±Ñло ÑазÑабоÑано пÑоÑÑое веб-пÑиложение, позволÑÑÑее лÑÐ±Ð¾Ð¼Ñ Ð¿Ð¾Ð»ÑзоваÑÐµÐ»Ñ Ð¿ÑимеÑиÑÑ Ð¸Ð½ÑеÑеÑÑÑÑий ÑÐ»ÐµÐ¼ÐµÐ½Ñ Ð³Ð°ÑдеÑоба. Virtual Try-on is a promising and fast-growing area of deep learning, which aims to transform the human appearance by adding new features (makeup, wigs), as well as changing existing ones (clothes). This paper considers the most popular and meaningful methods of virtual try-on clothes. The PF-AFN neural network was chosen as the main algorithm, which demonstrates the best results with the smallest input data requirements. During the work on the selected algorithm, many limitations of the existing implementation were found, so it was modified to take into account modern optimization methods, such as mixed precision learning and distributed data parallelization. At the same time the optimization did not affect the quality of the final result, as shown by the calculated metrics. To demonstrate the applicability of the method to industrial solutions, a simple web application was developed, allowing any user to try on the closet item of interest. |
Databáze: | OpenAIRE |
Externí odkaz: |