A Parallel Convolution and Decision Fusion-Based Flower Classification Method

Abstract

Flower classification is of great significance to the fields of plants, food, and medicine. However, due to the inherent inter-class similarity and intra-class differences of flowers, it is a difficult task to accurately classify them. To this end, this paper proposes a novel flower classification method that combines enhanced VGG16 (E-VGG16) with decision fusion. Firstly, facing the shortcomings of the VGG16, an enhanced E-VGG16 is proposed. E-VGG16 introduces a parallel convolution block designed in this paper on VGG16 combined with several other optimizations to improve the quality of extracted features. Secondly, considering the limited decision-making ability of a single E-VGG16 variant, parallel convolutional blocks are embedded in different positions of E-VGG16 to obtain multiple E-VGG16 variants. By introducing information entropy to fuse multiple E-VGG16 variants for decision-making, the classification accuracy is further improved. The experimental results on the Oxford Flower102 and Oxford Flower17 public datasets show that the classification accuracy of our method reaches 97.69% and 98.38%, respectively, which significantly outperforms the state-of-the-art methods.