Research And Implementation Of People Detection Base On Deeping Learning

Convolution neural network was put forward,which let image related industry get great breakthrough.However,there are still some problems in image research,such as low recognition accuracy and low recognition speed.For image object recognition,the detection accuracy is the most important,followed by the detection speed.At present,the effect of character detection in character images is general.The direct reason why the character detection accuracy is not high in the character image is that the missing detection rate of the character detection is relatively high,which leads to the low detection accuracy,which is the main problem to be solved.Secondly,for the speed of detection,the faster the speed of people detection requirements,the more scenarios applied,especially video processing.In this thesis,aiming at the current shortcomings of convolutional neural network,such as low accuracy of character detection,slow speed of network model detection,insufficient data volume to meet training requirements,unbalanced training data,etc.,the speed and accuracy of people detection in image are studied and realized.The main work of this thesis is as follows:(1)In order to solve the problem that the data quantity is small and the picture quality is not satisfied with the training demand,the method of data preprocessing and data enhancement is studied and designed.The method can detect the brightness and ambiguity of the detected image,filter out the under-exposed,over-exposed and fuzzy input pictures,and expand the data of the qualified image to produce more training data.(2)This paper designs and implements the speed raising scheme of convolution and batch standardization based on convolution neural network using Caffe framework.By improving the design of current convolution layer and batch standardization layer,the detection speed of detection model is accelerated.(3)Aiming at the problem of poor detection effect of convolution neural network in people detection,a feature cascade optimization pooling(FCOP)detection scheme is designed,which includes feature cross-layer hybrid detection.Multi-network merge detection and last layer split detection network three parts.By merging shallow features across deep convolution directly with the results of deep convolution,the scheme can detect multiple feature output layers,reduce feature loss,and optimize feature extraction method,and speed up the speed of feature extraction.(4)Finally,a practical test is carried out on the convolution batch standardization layer merging speed raising scheme and the feature cascade optimization pool scheme.The experimental results confirm the effectiveness of the character detection scheme proposed in this paper based on the depth learning in the actual scene.It can improve the precision and speed of people detection.