應用於行動裝置之基於深度學習的手繪運算放大器電路辨識與教學動畫生成系統開發

Abstract

隨著深度學習技術的發展，手繪文本和電路圖的識別取得了顯著進步。然而，針對手繪運算放大器 OPA 電路的研究仍較為有限。本研究提出了一種行動裝置專用的手繪運算放大器電路識別與動畫生成系統，旨在解決現有影像搜尋工具（如 Google Images）無法有效辨識手繪電路的問題。本系統採用最新的 YOLOv9t 目標檢測模型進行電子元件識別，相較於 YOLOv8n，在模型參數量（Parameters, Params）與每秒十億次浮點運算次數（Giga Floating Point Operations Per Second, GFLOPS）方面均有所提升。系統透過分析識別出的電子元件及其相對位置來確定電路類型，並允許使用者輸入元件參數，以 Manim 動畫引擎生成對應的輸出波形動畫，幫助學生直觀理解運算放大器電路特性。本研究構建了一個包含 1,199 張手繪運算放大器電路圖的資料集，並比較了 YOLOv8n 和 YOLOv9t 兩種物件偵測模型的辨識效能。實驗結果顯示， YOLOv9t 與 YOLOv8n 在 Precision 指標上均達到 99%，整體辨識校效能相當。然而 YOLOv9t 的參數量為 2.8M，較 YOLOv8n 的 3.0M 減少約 7%，在模型輕量化方面展現優勢；此外 YOLOv9t 的每秒十億次浮點運算次數為 11.7，遠高於 YOLOv8n 的 8.1，效能提升約 44.4%。顯示 YOLOv9t 更具運算效率，適合應用於本系統所需的即時辨識場景。除此之外，系統整合 LINE Bot 作為互動介面，使學生可直接透過行動裝置拍攝手繪電路圖，並即時獲得識別結果與動畫回饋。整體實驗結果顯示，本系統在電子工程教育領域具有潛在應用價值，未來將進一步透過使用者調查來優化互動設計與學習成效。With the development of deep learning technology, significant progress has been made in the recognition of handwritten text and circuit diagrams. However, research on handwritten operational amplifier (OPA) circuits remains relatively limited. This study proposes a handwriting recognition and animation generation system for OPA circuits specifically designed for mobile devices, aiming to address the issue that existing image search tools (such as Google Lens) cannot effectively recognize handwritten circuits. The system employs the latest YOLOv9t object detection model for electronic component recognition. Compared to YOLOv8n, it has improvements in both the number of model parameters and the Giga Floating Point Operations Per Second （GFLOPS）. The system determines the type of circuit by analyzing the identified electronic components and their relative positions, and it allows users to input component parameters to generate corresponding output waveform animations using the Manim animation engine, helping students intuitively understand the characteristics of OPA circuits.This research constructed a dataset containing 1,199 handwritten operational amplifier circuit diagrams and compared the recognition performance of two object detection models, YOLOv8n and YOLOv9t. The experimental results showed that both YOLOv9t and YOLOv8n achieved a Precision metric of 99%, indicating comparable overall recognition performance. However, YOLOv9t has a parameter count of 2.8M, which is about 7% less than YOLOv8n’s 3M, demonstrating an advantage in model lightweighting. Additionally, YOLOv9t’s GFLOPS is 11.7, significantly higher than YOLOv8n’s 8.1, representing an approximate performance improvement of 44.4%. This indicates that YOLOv9t is more computationally efficient and suitable for real-time recognition scenarios required by this system. Furthermore, the system integrates a LINE Bot as an interactive interface, allowing students to directly capture handwritten circuit diagrams using mobile devices and receive immediate recognition results and animation feedback. Overall, the experimental results indicate that this system has potential application value in the field of electronic engineering education, and future work will further optimize interactive design and learning outcomes through user surveys.