Research On Collaborative Target Segmentation Algorithm Based On Network Modulation

Object co-segmentation is a technology based on algorithms to segment objects of a common category in multiple related images.It focuses on the common features in the information of multiple pictures and mines common patterns among images of the same category.It has a wide range of applications in the field of computer vision.Traditional methods cluster different regions according to the brightness,color,or hand-designed edge features of the images,subject to limited feature representation capabilities,and complex images such as blurry and uneven lighting.It is difficult to achieve correct segmentation in the scene.With the rise of deep learning in recent years,object co-segmentation methods based on convolutional neural networks rely on massive amounts of data and powerful feature representation capabilities to greatly surpass traditional methods in performance.This paper uses deep convolutional neural network technology to make exploratory and innovative research on the object co-segmentation method based on network modulation,and has achieved the following results.This paper proposes a deep object co-segmentation network framework that combines spatial modulation and semantic modulation,with novel ideas and excellent performance.First,multi-scale image features are extracted through the backbone network as the input of the segmentation network.Secondly,a spatial modulator module is designed to capture the correlation between the feature descriptions of multiple images,and a set of rough masks learned through unsupervised learning It can locate the common foreground target position while suppressing the background.Considering that the co-targets actually have a common category,the semantic modulator is modeled as an image classification task.A cascaded second-order pooling module is proposed for image feature conversion,and the class prediction vector of co-target is obtained by supervised learning.Finally,the output of the two modulator modules and the multi-scale image features of the backbone network work together on the segmentation network in a modulated manner,focusing the features on the collaborative target to achieve segmentation.This method is end-to-end training and does not require any postprocessing,and has achieved superior segmentation accuracy on four public co-segmentation benchmark datasets.Among them,the most difficult PASCAL-VOC dataset is increased by 2.7percentage points compared with the same period method.Based on the previous work to achieve object co-segmentation,this paper further deepens the research to explore instance-level co-segmentation,that is,to obtain a segmentation result for each instance of the co-object.First,the co-peak search technology is used to find the peak points in the co-segmentation results,and then the peak back-propagation is used to obtain an instance-level response map on the original image,and multiple instance proposal regions of the image are obtained through an unsupervised method.Finally,the co-segmentation results,instance response maps and proposal regions are comprehensively scored and sorted,and the final segmentation result is selected as the highest score.This method does not require image instance segmentation labels for training,reducing the requirements for training data,and effectively expands the research depth from object co-segmentation to instance-level cosegmentation.This method greatly surpasses the benchmark method on the two public instancelevel co-segmentation benchmark datasets,and improves by 5 percentage points on the more difficult VOC12 dataset.