Hand Detector based on Efficient and Lighweight Convolutional Neural Network

Abstract

Hand detection and recognition topic has been studied since the last century and is particularly concerned with the development of machine learning today. Inspired by the benefits of a convolution neural network (CNN), this paper proposed an efficient and lightweight architecture to detect the location of hand in the images. This network is deployed with two main blocks which are the feature extraction and the detection block. The feature extraction block starts by convolution layers, CReLU (Concatenated Rectified Linear Unit) module, and max pooling layers alternately. After that, the six inception modules are used and final by four convolution layers. The detection block is constructed by three blocks of two-sibling convolution layers using for classification and regression. The experiment was trained on the combination of EgoHands and Hand dataset. As evaluation, the detector was tested on Egohands test dataset with the results achieved 93.32% of AP (Average Precision). In addition, the speed was tested in real-time by 33.87 fps (frames per second) on Intel Core I7-4770 CPU @ 3.40 GHz.