Design Simulation of Deep Learning Based Complex Flower Image Classification using Transfer Learning and Data Augmentation

Abstract

Flower classification is a challenging task due to the wide range of flower species which have similar shape, appearance or surrounding objects such as leaves and grass. In this paper, we propose a novel two-step deep learning classifier to distinguish flowers of a wide range of species. Firstly, the flower region is automatically segmented to allow localization of the minimum bounding box around it. The proposed flower segmentation approach is modeled as a binary classifier in a fully convolutional network framework. Secondly, we build a robust convolutional neural network classifier to distinguish the different flower types.

Not at all like basic item is classification, for example, recognizing felines from hounds, flower acknowledgment and classification a difficult assignment because of the wide scope of flower classes that offer comparable highlights: A few flowers from various kinds share comparative shading, shape and appearance. Besides, pictures of various flowers for the most part contain comparable encompassing articles, for example, leaves, grass, and so forth. There are in excess of 250000 known types of flowering plants arranged into around 350 families [1]. A wide scope of different applications including content-based picture recovery for flower portrayal furthermore, ordering [2], plants checking frameworks, gardening industry [3], live plant recognizable proof and instructive assets on flower scientific classification [4] rely upon fruitful flower classification. Manual classification is conceivable however tedious and dull to use with an enormous number of pictures and possibly incorrect in some classes flower particularly when the picture foundation is perplexing. Therefore, strong procedures of flower division, discovery and classification have extraordinary worth.

We propose novel steps during the training stage to ensure robust, accurate and real-time classification. We evaluate our method on three well known flower datasets. Our classification results exceed 98% on all datasets which is better than the state-of-the-art in this domain.