鞋類分割的分析YOLOv3在不同的背景圖像上

Abstract

機器學習技術已經廣泛應用於多種計算機視覺中，具備標準一致、高精度的優勢。本文旨在開發一種通過自動提取有意義的信息的方法，來對鞋類進行分割圖像。 文中提出已鞋類為對象的精確識別位置、標記類型的演算法。透過YOLOv3的網路框架實現，主要用於訓練鞋的特徵並驗證圖中未顯示的區塊。為證明有效性，採用兩不相同數據庫，共包含700個鞋類圖像。特別的是數據庫採用由乾淨背景及混雜背景的圖像組成的兩個數據庫，試圖已複雜的數據集模擬真實情境，使模型對於真實環境的答辯度提升。 在圖像上進行測試時，平均準確度分別為乾淨背景95％和復雜背景68％。另一方面，為使複雜背景的準確率上升，本實驗藉由乾淨背景數據集的權重轉移，來增強鞋的特徵。將來，這種方法可以進一步擴展，用作訓練其他數據集的基準，並結合更多樣的網絡體系、結構類型。

Machine learning technology has been widely used in a variety of computer vision applications, with the ability to produce consistent and highly accurate performance. This article aims to develop an automatic method for segmenting images of footwear by suggesting ways to extract meaningful information from it. The proposed algorithm generates the exact location of the footwear object and identifies the type of shoe. The method implemented in the experiment is called YOLOv3, which is mainly used to train the characteristics of shoes and verify invisible objects. To prove the validity, the two databases of the proposed method hold a total of 700 footwear images. In particular, the database consists of images with clean and complex backgrounds. In particular, complex datasets try to mimic reality, i.e.,It is more practical when implemented in practical applications. When tested on clean and tidy images, the average accuracy was 95% clean background and 68% complex background, respectively. On the other hand, to improve the accuracy of complex data sets, the model must first be made to learn the features of a clean data set. In the future, this method can be extended to further train other benchmark databases and combine network architecture types.