Research On Instance Segmentation Algorithm Based On ResNet-FPN

Scene instance segmentation is a research hotspot in the field of computer vision at this stage.It plays an important role in applications such as autonomous driving,video surveillance,intelligent robots,biomedicine,and virtual reality technology.The instance segmentation task can provide richer information than other tasks,and because of this,it is more challenging.When faced with the interference of complex backgrounds,the diversity of instances,and the mutual occlusion and truncation of objects,it is very important to design the detection process of instance segmentation to overcome the above difficulties in the inference process.The current instance segmentation related algorithms based on deep learning can be roughly divided into two categories,namely,bottom-up algorithms based on semantic segmentation and top-down algorithms based on target detection.The bottom-up algorithm first classifies the image pixel by pixel,and then further groups the pixels into different instances through clustering and other methods.T he topdown algorithm is to first detect instances,and then perform further pixel-by-pixel classification on the detected candidate regions.These two sets of algorithms have their own limitations when facing specific scenarios.By analyzing their respect ive characteristics,this paper proposes a method that combines the above two ideas to solve various challenges faced by instance segmentation.The bottom-up algorithm based on semantic segmentation can obtain larger receptive fields and higher resolution pixel-level features,while the top-down algorithm based on target detection can better perceive instance-level features.In order to make the model possess the respective advantages of the above two algorithms,adapt to different scenarios,and obtain pixel-level and instance-level features at the same time,this paper proposes an information fusion scheme.The designed information fusion module can synthesize the information obtained by the two algorithms,and the complementarity is insufficient to genera te the final prediction result.Because the structure of multitasking is too complicated to be optimized,this paper proposes an instance segmentation network that can perform end-to-end training,so that each process shares part of the structure and can r un in parallel,thereby greatly improving the overall efficiency of the model.In addition,in order to make the model more adaptable to different scenarios,this paper introduces an unknown category perception mechanism and a deformable perception mechani sm,so that the network can avoid mispredictions as much as possible,and increase the network's adaptability to various complex deformations.In this paper,a series of experiments are conducted on the COCO and Cityscapes datasets to verify the effectiveness of this model.Compared with the benchmark model,the average accuracy of the proposed method on the COCO dataset and Cityscapes datasets has been greatly improved.