Anchor-free Based Object Detection And Instance Segmentation Methods

As one of the most essential means of realizing artificial intelligence,deep neural networks have been widely used in recent years.Among them,convolutional neural networks have boosted the development of the computer vision field significantly.As the research becomes more and more intensive,the computer can conduct image processing from instance-level to pixel-level.Object detection and instance segmentation are the most basic and difficult tasks,the huge model and complex algorithm restrict them from playing important roles in real-world manufactury.Based on the anchor-free method,this thesis proposes a more efficient and accurate object detection and instance segmentation method.While greatly reducing the hyper-parameters and training-parameters of the model,by encoding the bounding boxes as vectors for regression,a novel vector-encoded loss is proposed,which can improve the accuracy on open datasets markedly.Moreover,a new activation function is proposed to help the network converge to the global optimal value more quickly and reduce the training time.In order to demonstrate the effectiveness of the proposed algorithm,experiments were carried out on multiple open datasets containing image classification,object detection and instance segmentation.The results show that the method of this thesis can get more accurate bounding boxes and segmentation mask at a faster speed and with no extra training parameters.Firstly,the design idea of the activation function is studied.Since the activation function is a nonlinear function,it has the ability to help the neural network to fit a nonlinear mapping from input to output.However,the computational complexity of the activation function,the gradient update amplitude,and the effect on the convergence should be taken into account when the nonlinearity is realized.In this paper,the advantages of piecew ise activation function and non-piecewise activation function are integrated,and the disadvantages of both are abandoned.A continuous non-piecewise activation function is designed,which can significantly improve the accuracy and robustness of the network while greatly reducing the amount of computation.Secondly,in view of the regression process in the object detection task based on anchor-free methods,new regression algorithms based on the global and the center distance is designed.At the same time,considering the fact that the direction is required to supervise the regression of the bounding box to the ground truth box during the regression process,the vector information of the direction is added.Finally,the designed vector encoded loss can provide supervision information for bounding boxes which have an Io U of 0 or those within the ground truth boxes.Through the experiments on the dataset of object detection,it is proved that the convergence speed and the final result of the proposed algorithms are significantly better than other previous algorithms.Then,aiming at the problem that the structure of the anchor-free object detection network has not been fully optimized,an improved object detection network combined with the continuous non-piecewise activation function and the regression loss based on encoded vectors is proposed.The improved structure includes the head detection module and the loss function of the regression branch.After verifying the effectiveness of the network structure in the common dataset,experiments were also carried out on more difficult remote sensing datasets to test the effect of the network.The experimental results show that the improved anchor-free object detection algorithm can keep robust in different datasets,meanwhile can effectively improve the detection accuracy without increasing the training and testing time.Finally,the problems in the structural design of the anchor-free instance segmentation algorithm is analyzed.Because one-stage instance segmentation algorithms can not use the region proposal network to provide the information for reweighting the segmentation of the region of interest,the effect of the spatial attention guidance module is not significant enough.Considering that the segmentation part of the network uses the region of interest itself to reweight,the channel self-attention mechanism module is used to guide the segmentation.The experimental results show that using the parellel self-attention mechanism is more significant than using the spatial attention module itself.