Research On Target Detection Based On Deep Learning

As a classic topic in the field of image processing and computer vision,target detection has been widely used in traffic monitoring,image retrieval and human-computer interaction.It is designed to detect objects that people are interested in,in a static image(or dynamic video).In traditional target detection algorithms,feature extraction and classification decision are performed separately,and the requirement of feature selection is more strict.It is difficult to get ideal effect when facing complex scenes.Since Professor Hinton puts forward the theory of depth learning,more and more research scholars try to adopt the depth learning concept to solve the problem of target detection,and put forward different models.Different model applications are not the same,and convolutional neural networks are usually used to deal with the problem of target detection.Compared with the traditional target detection algorithm,the feature extraction and pattern classification in the convolutional neural network are carried out in parallel,and the complex scene can be handled better with the increase of the number of layers,but it has bad constraint on target edge.On the basis of this,the traditional algorithm and the convolution neural network are studied deeply,and a target detection algorithm combining traditional algorithm and convolutional neural network is achieved.The main work and innovation of this paper are as follows:(1)The traditional target detection algorithm generally uses the way of rectangular to get the approximate area of the target,and our demand is as much as possible to obtain the target's edge contour.This paper achieves an improved target detection algorithm based on active contour models so that contours are as close as possible to the target.(2)In order to solve problems that the traditional algorithm requires manual design of image features and different scene models are unstable and the convolutional neural network segmentation is not accurate and the constraints between the adjacent pixels are missing,this paper will combine the traditional algorithm and convolutional neural network,using convolutional neural network for "high-level" images feature extraction,and super pixel for extracting "low level" images characteristics.It can adapt to different complex scenarios,and obtain accurate target edge.Extensive experiments are carried out in the the Summer Palace database.Through the results we can see that the our algorithm for target detection and extraction can extract the target accurately,and the edge constraint of the object is very strong.(3)The main innovation points:On the use of super pixel features extraction removing duplicate features,can reduce feature redundancy and reduce the dimension of the feature;At the same time,we should change VGGNet(Visual Geometry Group Network)network into GoogleNet network with a faster convergence,to improve the speed of the algorithm.