基於階層式時序記憶的多角度手勢辨識方法

Multi-Angle Hand Posture Recognition Based on Hierarchical Temporal Memory

黃雅軒 ； 王勻駿

《Journal of Information Technology and Applications(資訊科技與應用期刊)》 6卷4期 (2012/12) Pp. 151-161

https://doi.org/10.6302/JITA.201212_6(4).0001

摘要

在圖形辨識的各種應用中，角度變化一直都是影響辨識效果的主要因素。為了提高角度變化的容忍能力，本論文使用了階層式時序記憶（Hierarchical Temporal Memory，HTM）演算法。此演算法的特性乃是藉由時間性的概念，將時間上的連續變化影像特徵進行歸納，構成「不變性特徵」，因而可以克服角度變化的影響。本論文所提出的多角度手勢辨識方法，主要包含手勢影像前處理和手勢辨識兩個模組。在手勢影像前處理模組方面，輸入影像個別經由膚色偵測、背景分離和邊緣偵測等處理；然後，將上述三種處理的輸出結果，藉由線性權重來正確的取出手勢區域；接著，判斷此手勢區域是否擁有手的前臂部分，若存在前臂，則進行前臂分離處理，只保留手掌的部分；再將處理後的影像正規化，形成具有固定大小的手掌影像。在手勢辨識模組方面，它將正規化後的手掌影像以HTM演算法來進行各種手勢模型的學習與辨識。經由實驗證明，針對相同的測試資料庫，本論文提出的方法對於剪刀、石頭和布等三種手勢，可達到89.1%的高辨識率，而相較於Adaboost和SVM兩個演算法的78.1%和79.9%辨識率，本論文的方法的確可以達到更好的辨識效果。

關鍵字

階層式時序記憶 ； 多角度手勢辨識

並列摘要

In the field of pattern recognition, angle variation plays an important role in producing effective recognition results. To overcome the angle variation problems, this paper adopts the Hierarchical Temporal Memory (HTM) algorithm. Based on the inherent property of the HTM algorithm which applies temporal information to organize the continuous change in time of image features in constructing their respective ”invariant features”, the effect of angle variation can be overcome. The proposed ”Multi-Angle Posture Recognition Based on HTM” are consists two modules of ”Hand Posture Image Preprocessing” and ”Hand Posture Recognition”. At ”Hand Posture Image Preprocessing” module, the input images will then be individually processed by skin detection, background segmentation, and edge detection. The processed results are further combined with a linear weighting method to acquire the correct hand posture region. If a forearm exists, a forearm segmentation step will be executed and keep only the part of palm. Then, the processed palm images will be normalizing to fixed size. At ”Hand Posture Recognition” module, the normalized images are forwarded to HTM for learning and recognizing of varied hand posture models. The experiment results show that when using the same testing database, the proposed method can achieve 89.1% high recognition rate with three kinds of hand posture, such as scissors, stone and paper, thereby resulting in better performance then both Adaboost and SVM learning algorithms that individually only achieved 78.1% and 79.9% recognition rate.

並列關鍵字

Hierarchical Temporal Memory ； Multi-angle hand posture recognition