Research On Image Target Detection Algorithm Based On Deep Learning

The design of target detection is based on a wide range of theories,including pattern recognition,artificial intelligence,image processing and so on.It has a large number of applications in many fields,such as video surveillance,human-computer interaction and so on.With the rise of depth learning,the depth learning module is introduced into target detection,which makes the algorithm work stably in complex scenes.The depth neural network abstracts the feature of the target object through multi-layer network and the image under test from the shallow network to the deep network,so the information extracted is more comprehensive and the detection effect is more effective.Depth learning and its application in target detection are studied in this thesis.The following improvements based on the YOLOv3 model are made.Firstly,the YOLOv3 model uses the classical K-means clustering algorithm to determine the anchor location.However,the algorithm is random in selecting the initial position of the anchor,which causes the contingency and affects the clustering results.An improved kernel K-means clustering algorithm based on sampling is proposed in this thesis,which combines kernel function and sampling on the basis of classical K-means clustering algorithm to determine the initial position of anchor.The experimental results based on MNIST show that the clustering effect of the improved algorithm increases from 94.72% to 95.58%,and the location of the initial anchor is more accurate.Then,in order to make the YOLOv3 model more sensitive to small and dense targets,a kind of deep convolution neural network combined with inception structure is proposed.Dark_inception network uses inception structure to replace a group of convolutional layers in DarkNet-53.Dark53_inception network adds inception structure after multi-scale fusion of YOLOv3 model.The two models not only widen the width of the network,but also reduce the calculation of the parameters.Experimental result based on Open dataset VOC 2007 shows that the mAP of the improved network is 1.76 and 2.07 times higher than that of the unimproved YOLOv3 network,and the detection accuracy of the improved network is improved to some extent.Finally,aiming at the shortcoming that YOLOv3 model is not accurate to detect small targets,this paper proposes that adding one scale and increasing the number of three anchor on the basis of the fusion of three scales of the basic model and the Dark_inception model respectively.The shallower network is fully considered the features.Experimental results show that the improved network is more accurate than the unimproved network prediction box,and the accuracy is increased by 1.54 times.