Research On High Precision And Fast Target Detection Method Based On Deep Learning

Further the research on deep learning algorithms goes,target detection algorithm based on deep learning has become one of the important research topics in the field of computer vision,plays an important part in many occasions: intelligent control,intelligent transportation,military fields,etc..After comparing different types of convolutional neural network(CNN),it is clear that the accuracy and operating efficiency of existing target detection algorithms need further improvement: in videos and pictures,influenced by distance,insufficient light or shadow and other factors,some target images have small pixels,which are difficult to be detected by some target detection methods.Aiming at the above problems,the improved Faster R-CNN algorithm is adopted in this paper to detect the target.Based on SOFTNMS method,Focal Loss and Loss function combination method and Light Tail Faster R-CNN method were proposed to solve the above problems The main content of this paper is as follows:1.In order to reduce the impact of missed and rechecked targets on target detection,and achieve accuracy improvement.Introduce SoftNMS into Faster R-CNN to replace the original NMS structure so as to optimize the region generation network.The detection accuracy is improved by 4.3% compared with the original algorithm.2.As for the problem of sample imbalance,the accuracy of target detection is too low.Combine Focal loss function with loss function,then the accuracy of classification and recognition will get further improvement.The experimental results prove that the detection accuracy of the improved Faster R-CNN algorithm is improved by 6.2% compared with the original algorithm.3.In the target detection task,the network operation efficiency determines the actual effect of the real scene application.On the basis of Faster R-CNN model,this paper proposes a Light Tail Faster R-CNN network architecture,aiming at the impact on the speed of target detection caused by the complex and redundant network structure of Faster R-CNN.Light Tail Faster RCNN replaces the full connection layer in the network structure with a 1×1 convolution layer so as to achieve the speed improvement.Experiments on PASCAL VOC datasets prove that within the range of 0.62% accuracy loss.,Light tail Faster R-CNN algorithm can effectively improve the speed by 53%.4.In the actual face detection task,it is difficult to detect small pixel targets due to the interference of image Angle,ambient light,shadow and other factors.In this chapter,the improved algorithm mentioned above is applied to face detection.The experimental results show that in the simple data set,the high-precision Faster R-CNN target detection method based on SOFTNMS is 6.6% higher than the original algorithm Faster R-CNN,and the high-speed Faster R-CNN target detection algorithm based on Light Tail increases the detection speed by51% when the accuracy decreases by 1.56%.In complex scenes,the missed detection rate and error detection rate of the improved algorithm are reduced by 25.6% and 50% respectively,which improves the accuracy of the algorithm on the whole and verifies the effectiveness of the improved method.