Instance Segmentation Method Based On One-Stage Object Detection Model

The task of object detection and instance segmentation are to distinguish each object in the image.They have certain similarities.And so many works complete instance segmentation task based on object detection.The two-stage object detection model can complete the task of instance segmentation by simple extension,but it is difficult to do well in one-stage model.This work is to design an instance segmentation extension method based on a one-stage object detection model,so that this simple and universal extension method can obtain high-quality instance segmentation results while maintaining the detection performance.The benchmark object detection model identifies objects of different sizes at each layer of the feature pyramid(object streams),and expresses each attribute(category,detection box)of the image in the form of point objects,but it is difficult to express the segmentation mask of an object in the same way.For this reason,the extension method in this paper firstly expands a fully convolutional segmentation stream at the bottom layer(P3)of FPN.The segmentation stream outputs a group of segmentation foundations,and at the same time adds a group of combination coefficients corresponding to the segmentation foundations to the object stream.An image generates a group of segmentation foundations through the network,and the combination coefficients generated by the object stream is responsible for distinguishing different objects,and the linear combination of this two tensors finally obtains the instance segmentation mask of the object.The problem of inconsistent convergence of object stream and segmentation stream will occur in the training of the expanded model.This paper proposes to apply different learning rate to different convolution streams to solve.At the end of the extended model is a post-processing reorganization method,which improves the processing of the segmentation mask based on the basis of the original NMS algorithm and improves the accuracy of instance segmentation without affecting the original object detection performance.The linear combination process of segmentation foundation and combination coefficient involves three activation functions,and different combinations of activation functions will lead to different segmentation foundation results.In this paper,the effects of different combination modes are analyzed through diagrams,and a group of combination modes with the best segmentation accuracy result is obtained: the combination coefficient is activated by the softmax function,the segmentation foundation is activated by the sigmoid function,and no activation is required after linear combination.However,this combination method is still difficult to deal with the situation that similar objects have overlapping detection boxes.This paper proposes a union region loss strategy to solve this problem.This strategy enables the model to learn the many-to-many relationship between point objects and segmentation foundations,and really enables two objects with overlapping relationship are “decomposed” into two different segmentation foundations.Finally,this paper also proposes a cross-attention module,which is a connection module connecting the C2 and P3 layer of the network,which can effectively improve the segmentation stream.In this paper,during the expansion of the one-stage object detection model,several improvement methods are proposed.These improvement methods have the same generality as the extension based on combination method,which can be added to different models and bring improvement.