影像辨識技術應用於拆解廢電子電器計數之可行性評估計畫
| 中文摘要 | 本計畫蒐集國內外運用影像辨識應用於廢棄物回收處理之技術發展與案例，主要應用可分為辨識分類與物件偵測兩大方向，並檢視現行稽核認證作業流程六大程序，評估影像辨識技術於廢棄物拆解處理作業較具應用可行性。 依據本計畫之影像辨識模型試驗結果，並考量AI智慧影像辨識需求，評估既有CCTV影像不適合用於AI辨識與計數，為取得供機器深度學習素材，本計畫至5家受補貼機構蒐集LCD電視機、冷暖氣機、電冰箱以及洗衣機各品項高解析度約2,500小時之拆解處理影像，完成廢物品各1,000張影像標記作業，建置廢物品影像辨識模型學習資料集。以YOLOv4物件偵測演算法為核心建置影像辨識模型，模型平均精準率(mAP)達97.7%，利用實廠影像進行測試，模型召回率與精準率達90%，顯示影像辨識技術可偵測與辨識複雜外觀的廢電子電器，再結合物件追蹤演算法與計數線輔助邏輯，優先以流水線拆解作業模式的廢電冰箱為試驗標的，試驗結果發現單廠單日影像辨識計數量與稽核認證量差異為2台，顯示於特定條件下，影像辨識技術初步已具計數可行性。為提升影像辨識技術之應用可行性，本計畫已研擬技術面與行政面之配套措施，辦理一場次專家諮商會議精進本計畫研議之內容，計畫執行期間，並配合參加一場次專家交流會。 | ||
|---|---|---|---|
| 中文關鍵字 | 影像辨識技術、廢電子電器、稽核認證 | ||
基本資訊
| 專案計畫編號 | 經費年度 | 110 | 計畫經費 | 3680 千元 | |
|---|---|---|---|---|---|
| 專案開始日期 | 2021/01/27 | 專案結束日期 | 2021/12/31 | 專案主持人 | 黃欣栩 |
| 主辦單位 | 回收基管會 | 承辦人 | 楊錫桂 | 執行單位 | 財團法人中興工程顧問社 |
成果下載
| 類型 | 檔名 | 檔案大小 | 說明 |
|---|---|---|---|
| 期末報告 | 影像辨識技術應用於拆解廢電子電器計數之可行性評估計畫成果報告本文(公開版).pdf | 16MB |
Feasibility study for the application of image recognition technology on auditing waste appliances treatment
| 英文摘要 | This project collects domestic and foreign cases of using image recognition to assist waste recycling and processing, and grasps the core technology and application aspects; reviews the current six major procedures of the audit and certification process, and uses image information to identify and count waste products and control the flow of waste products. Purpose, the evaluation technology is suitable for the dismantling process, and can combine the relevant information flow with the existing forms, which can reduce the workload of the work. This project examines the CCTV images of the existing dismantling and processing areas in China. Considering the image recognition conditions, it has been assessed that the existing CCTV images are not suitable for AI identification and counting. However, some of the information can be used to monitor the flow of waste products through angle and position adjustment.This project collects high-resolution images for dismantling and processing of LCD TVs, air-conditioners, refrigerators, and washing machines from 5 subsidized organizations, and completes 1,000 images of waste products and 500 images of spindle components. Set the image recognition model learning data set.This project uses the YOLOv4 object detection algorithm as the core to build an image recognition model. The average accuracy rate (mAP) of the model is 97.7%, and then verified with real factory images. The model recall and precision rate reach 90%, showing that the image recognition technology can detect Measure and identify waste electronic and electrical products with complex materials and appearances.Combining an image model with certain recognition capabilities with an object tracking algorithm and counting line auxiliary logic to construct a model counting function. After evaluating the on-site operation mode, the dismantling of waste refrigerators in the assembly line is more feasible for technical introduction. Therefore, the waste refrigerators are the test targets for waste counting. After the AI misjudgment problems are summarized and adjusted, the model counts and audits. The difference in the amount of certification is 2 sets and the counting is feasible in certain condition. The evaluation result of the recognition rate of this project shows that the image recognition technology can identify, classify, track and count the number of waste products, combined with the aforementioned application-oriented supporting measures for the development of technical and administrative aspects, including the identification and counting principles, and the application of software and hardware. Specifications and identification information transmission structure, etc. | ||
|---|---|---|---|
| 英文關鍵字 | Image recognition technology, Waste appliances, Auditing and certification | ||