Research On Image Object Detection And Segmentation Methods Based On Deep Learning

Object detection and instance segmentation are very important research topics in the field of artificial intelligence,and they have significant practical applications in industry.Object detection and instance segmentation are usually linked and they complement each other.The improved accuracy of object detection will help the model to determine the region to be segmented,and the segmentation results will help the model to fine-tune the bounding boxes.In this paper,we investigate the problems in object detection and instance segmentation in the present stage and accomplish the following work.In order to solve the problem that the complexity and accuracy of neck structure in target detection are difficult to reach a balance,this paper designs an efficient and decoupled neck structure Shared-FPN.Neck structure is very helpful for performing multi-scale information fusion.Earlier neck structures were difficult to enhance significant detection effects for the model due to their simple structure.In recent years,the neck structure incorporates the technique of neural network architecture search,which leads to a very high model complexity.Shared-FPN does not increase the number of model parameters and GFLOPs,and can be easily ported to any detection model.It improves on PAFPN by designing a convolution module based on transposed convolution,a shared convolution and a Spatial Pyramid Pooling-Fast Down-sampling(SPPFD)module.Shared-FPN is able to achieve excellent results in the detection of small,medium and large size objects in both VOC 2012 and COCO 2017 datasets.In the VOC 2012 dataset with mostly large objects,the m AP of Shared-FPN is improved by 11.2% compared to FPN.In the COCO dataset with mostly small objects,the m AP of Shared-FPN is improved by 7.8% compared to FPN and by 7.1% compared to PAFPN.Shared-FPN can help models improve performance in tasks that require the use of object detection in industry,such as autonomous driving,medical detection,etc.In order to further improve the robustness and the edge segmentation effect of the existing segmentation models,a new instance segmentation model EEMask(Edge Enhanced Mask)is proposed in this paper.The speed of the single-stage instance segmentation model is fast,but the accuracy is difficult to surpass that of the two-stage model;the accuracy of the two-stage model is high,but the inference time is difficult to accelerate.Both types of models perform poorly in segmentation at instance edges.EEMask cleverly handles high-level features and low-level features,integrates the advantages of different scale feature maps,calculates the relevance between high-level features based on distance and feature values,and designs edge enhancement layers for low-level features to enhance the model’s ability to perceive the edges of instances.The EEMask model is evaluated on the COCO 2017 dataset,and it has an inference speed of 11% faster than Blend Mask,m AP improvement of 3.6% compared to Blend Mask,and 7.3% compared to Mask RCNN.In industry,instance edge precision is required for tasks where EEMask can help to better accomplish the task,such as robotic arm grasping,automatic obstacle avoidance,etc.