Image Segmentation Based on Convolutional R-CNN

With the development of artificial intelligence, probabilistic graphical models, and machine learning, computer vision has made significant breakthroughs, making image instance segmentation a new challenging task in computer vision following semantic segmentation and object detection. Compared to traditional computer vision tasks, instance segmentation not only requires distinguishing different objects in an image but also differentiating between instances of these objects, making the task highly comprehensive. Many algorithms have been developed to solve instance segmentation tasks, among which the Mask Region-based Convolutional Neural Network (Mask R-CNN) algorithm is relatively mature. However, there is still room for improvement in Mask R-CNN, especially in terms of segmentation boundaries where significant errors exist. To solve the issue of blurred instance segmentation boundaries, this paper explores a method combining Mask R-CNN with Conditional Random Fields. This method employs convolutional conditional random fields and fully connected conditional random fields in the mask branch to optimize the further segmentation of candidate regions, resulting in more precise boundaries. Experimental results show that compared to the original algorithm, these methods increase the mAP value by 0.6% and 3%, respectively. In the comparison of segmentation results for targets of different scales, the improvements with Mask R-CNN-ConvCRF are particularly notable for medium and large-scale targets, with increases of 4% and 5.1%, respectively.