應用於低成本單板電腦之有效人臉偵測設計

Abstract

人臉偵測(Face Detection)主要是在任意圖像中找到人臉的位置和大小的技術。早期人臉偵測演算法受到了燈光及遮蔽物影響其偵測準確率，隨著深度神經網路的發展已可以解決人臉偵測遇到的多數問題，並且獲得穩定的準確率，但是深度神經網路往往需要強大的硬體設備才能完成，昂貴的硬體設備並不符合許多實務的需求。單板電腦雖然價格低廉能夠大幅降低人臉偵測的門檻，但是單板電腦僅提供入門級的硬體設備，硬體的計算能力遠遠不及一般的電腦。 本論文的研究目的是提出應用於低成本的單板電腦之有效人臉偵測網路設計，以降低人臉偵測硬體設備需求，減少計算量，並且維持人臉偵測的準確率。設計的內容主要包含使用模型權重量化，預選框設計、訓練集篩選，最後透過樹莓派3B來實現人臉偵測。實驗結果顯示本論文所提出的方法，除了能夠有效降低偵測時間外，並能增加2.7%的平均準確率，實際偵測一張影像的時間只需要0.3秒，達到近似及時人臉偵測的目的，並且成功地利用低成本的設備準確且快速的偵測人臉的所在位置。

Face detection is a computer technology that identifies human faces in digital images. Light and facial occlusion affect the accuracy of face detection. Most face detection problems can be solved by deep learning model to get high accuracy. However, the use of deep learning model generally requires expensive and powerful devices. Single board computer is a cheaper device for face detection. Unfortunately, Single board computer is a less powerful device than general computers. In this study, we proposed a face detection model design for low-cost single board computers so that it can reduce computing power and increase detection accuracy. The detection model design uses the quantized model, a set of modified anchor boxes, the dataset analysis. The overall design is implemented on the Raspberry Pi 3B. The experimental results verify that our method can increases the average accuracy by 2.7 %, and its cost time requires only 0.3sec. The proposed method can effectively detect faces with the use of the low-cost device.