Research On High Performance Gesture Detection Algorithm Based On Improved Deep Learning Architecture

Gesture detection plays an important role in computer vision.Gesture detection technology includes the regression and classification functions of gestures,the regression function can locate the position of human hand in the image,and the classification function can identify which category it belongs to from different gestures.Since human hands only account for a small part of the human body,some gesture targets in the whole picture are small gesture targets,and the current gesture detection algorithms often fail to achieve the expected goals at the same time in terms of detection accuracy and speed.With the rapid development of deep learning technology,target detection algorithms based on deep learning are becoming more and more mature.This article focuses on the shortcomings of the existing gesture detection algorithm to carry out the following work:In order to solve the problem of low detection accuracy for small gesture targets in Faster R-CNN gesture detection,an improved gesture detection algorithm based on attention mechanism and multi-scale fusion is proposed in this paper.Firstly,the attention module is added to the VGG16 feature extraction network in the original algorithm to fully extract the gesture features in the image.Then,the scale fusion technology is used to fuse the three feature images of different scales of the feature extraction network to increase the detection ability of small target gestures.Experimental results show that the improved Faster R-CNN algorithm achieves an average accuracy of 98.84%on ASL gesture data set and 93.5%on self-made gesture data set.The results show that the improved method achieves better detection accuracy.In view of the imbalance of positive and negative samples in gesture detection of YOLO v4 and the low accuracy caused by the difficulty in detecting small targets,this paper proposes a YOLO v4 gesture detection algorithm with improved loss function.Firstly,use K-means++algorithm clustering to obtain 7 Anchor Boxes suitable for gesture data sets to improve the convergence speed of the algorithm.Then,DY-Re LU activation function is used to improve the backbone network of the original algorithm.Finally,a loss function with penalty terms is proposed to balance the imbalance between positive and negative samples and increase the detection ability of small gesture targets.Experimental results show that the improved YOLO v4 algorithm has an average accuracy rate of 98.86%on the ASL gesture data set,a detection speed of 56.51f·s^-1,a detection accuracy of 95.1%on the self-made gesture data set,and a detection speed of 50.24f·s^-1.The results show that the improved method can improve the accuracy of gesture detection while ensuring the real-time performance.