Constrained Convolutional Neural Network And Its Application In Target Direction Estimation

At present,in the field of computer vision,the mainstream convolutional neural network(CNN)algorithms focus on the target recognition and positioning,and most of them use the axis-aligned bounding box to position targets.However,in order to make a deeper understanding of images,it is necessary to obtain the direction information of targets.Many target direction estimation researches based on CNN directly regress and predict the direction angle 0 which indicates the direction of the target.However,there is a large numerical jump around 0° and 360°,which may introduce errors into the network.Therefore,a new method for image target direction estimation is proposed in this paper.The convolutional neural network is used to regress the two monotonous and continuous unit directional components are(sin0,cos0)of the target direction angle 0.Because there is a function constraint that the sum of squares equals to 1 between the two components of the direction angle,that is,a new problem—the output constraint problem—is introduced into the neural network.In fact,many applications have output constraint problems.In order to solve this kind of problem,this paper analyzes different situations with output constraints,and divides them into weak constraints with output range constraints on the value domain and strong constraints with exact function constraints among output components.For each kind of constraints,general methods are proposed to reconstruct the traditional convolutional neural networks and then construct constrained convolutional neural networks(CCNN)to solve this kind of problem with output constraints and improve the fitness of the network models.There are output conversion method andadding constraint errors in the Loss layer.The output conversion method converts the outputs to meet the required constraints by using suitable conversion functions.Adding constraint errors in the Loss layer directly incorporates the constraint errors into the error back propagation in the neural network,and participates in the correction and calculation of the weight value,which aims at creating a weight parameter matrix that contains constraint information.In the target direction estimation network of which the outputs are designed as directional components,the proposed two methods are specifically used to construct the constrained convolutional neural networks.Discuss the specific network architecture design,the unitized transfer function design,and the modified error back propagation derivation,etc.A detailed scheme is given,and a comparison experiment of the two methods and the traditional method that does not consider the output constraints is compared,including experiments on single-type targets and multiple-type of targets.The single-type target comparison experiment shows that,the two kind of CCNN proposed in the paper are superior to the traditional CNN which does not consider output constraints,in improving the accuracy of directional component estimation and reducing the constraint error.The CCNN by using output conversion method can directly improve the accuracy of the direction estimation and completely resolve the problem of constraint errors.Increasing the method The CCNN by adding constraint errors can reduce the constraint errors of directional components more stably without affecting the decline speed and magnitude of the original output loss,and then improve the overall estimation accuracy.Adding constraint errors method is more universal than the output conversion method.It can also be used when the output constraint relationship is too complex and it is impossible to find a suitable output conversion function.When simultaneously performing target recognition and direction estimation for multiple types of targets,the two CCNNs have no influence on the original target recognition performance,and the same rules are exhibited in the direction estimation as in the single type.Therefore,for the practical problems with output constraints,if there is an appropriate conversion function that fully meets the constraint problem,the CCNN constructed by the output conversion method can reduced the output errors greatly r,and improve the estimation accuracy,and directly avoid the constraint errors.The CCNN by adding constraint error method is more universal,which can be applied to all constraints,can reduce the output constraint error more stably,and then affect the output estimation error to a certain extent,and also improve the estimation accuracy.