應用深度學習之即時蔬菜辨識系統整合價格數據之研究

市場上已經有用於識別植物，花卉、水果和蔬菜的App軟體，這對於識別植物、水果和蔬菜非常有教育意義。首先，辨識隨著手機裝置普及，手機裝置上的查詢數量也增加了，使用著通常在網路上搜尋料，但可能並不是真正你想的東西在網路上找到資料。本文提出將深度學習應用於即時蔬菜識別系統，並結合大數據蔬菜行情分析與預測，使用者透過行動裝置的相機對要識別的蔬菜拍照，再以深度學習模型辨識蔬菜，連線即時查詢蔬菜的種類和介紹健康信息，並向使用著提供蔬菜的歷史價格變化，可以進行比較和查詢蔬菜浮動價格。本文提出一套引用重新訓練 Google inception v3 的權重模型做為即時圖像分類蔬菜種類之辨識系統，也實作其它兩個方法來做比較，使用Resnet50和LeNet5兩種機器學習模型，進行圖片辨識學習。先將收集到的蔬菜照片做機器學習訓練，訓練好的機器學習模型放置於雲端伺服主機，使用者以手機相機拍照蔬菜，系統即可自動連結雲端進行深度學習辨識，辨識出蔬菜後再彙整健康資訊跟菜價回傳顯示於手機或網頁上。經過實驗驗證後，本系統辨識率為 95%以上。

關鍵字

辨識系統； Google inception v3 ；深度學習

並列摘要

There are already App software for identifying plants, flowers, fruits and vegetables on the market, which is very educational for identifying plants, fruits and vegetables. First of all, with the popularization of mobile devices, the number of inquiries on mobile devices has also increased. It is usually used to search for materials on the Internet, but it may not be what you want to find data on the Internet. This article proposes to apply deep learning to the real-time vegetable recognition system, combined with big data vegetable market analysis and prediction, users take photos of the vegetables to be recognized through the camera of the mobile device, then use the deep learning model to identify the vegetables, and connect to the real-time query of the vegetables. Types and introduction of health information, and the historical price changes of the vegetables used to provide comparisons and query vegetable floating prices. This paper proposes a set of weight models that use Google inception v3 as a recognition system for real-time image classification of vegetable types. It also implements the other two methods for comparison, using two machine learning models, Resnet50 and LeNet5, to perform picture recognition learning. The collected photos of vegetables are used for deep learning training first, and the trained deep learning model is placed on the cloud server host. Users can take pictures of vegetables with a mobile phone camera, and the system can automatically connect to the cloud for deep learning identification. Information and food price returns are displayed on the phone. After experimental verification, the system identified as more than 95%.