Research On Object Detection And Instance Segmentation Algorithms Based On Deep Learning

In recent years,deep learning theories and applications have made great progress in many fields,and its effects in text,speech,image and other fields are significantly better than traditional methods.Among them,object detection and instance segmentation are the active directions of long-term research in the field of computer vision.It has great application value in many fields,such as autonomous driving,face recognition,agricultural product classification,etc.This mainly depends on the deep neural network,and it plays a key role in these fields.The technology of deep neural network comes from the bionics of cells,imitating the way human obtain visual information.At present,most of the commonly used research methods in the field of object detection and instance segmentation are methods based on deep neural networks,and convolutional neural networks are one of them.The data set mainly used for network training is a manually labeled data set,which contains a large number of labeled data.When in use,the artificially labeled data is input into the deep neural network for training.The current object detection model and instance segmentation model have problems such as imperfect feature extraction,difficult feature fusion,inaccurate frame position prediction,and inaccurate instance segmentation.Improving the accuracy of object detection and instance segmentation has always been the goal of model design.Therefore,this paper proposes a "one-stage" object detection algorithm with fast detection speed and high accuracy and an instance segmentation algorithm with high segmentation accuracy for the above problems.The main research contents of this paper are as follows:Firstly,for perfect feature extraction of the object detection model,this paper proposes an improved backbone network.To improve the feature extraction ability of the backbone network,this paper constructs a cross-path convolutional aggregation network based on the attention mechanism.The use of ResNeXt in the convolutional neural network reduces the number of parameters of the model and widens the dimension of the model,ensuring the network feature extraction capabilities.The attention mechanism's introduction makes the network containing suitable weight distribution,paying attention to the structure and location information of the target.Secondly,for the feature fusion of the object detection model and the prediction of the target's category and position,the neck network between the head networks is improved.A composite feature fusion method is used,including top-down and bottom-up connection methods,and the Spatial Pyramid Pooling is used to fix target's feature size.Experiments have shown that reasonable fusion of features at different levels can effectively improve the object detection's accuracy.Aiming at the problem of target category and location prediction accuracy,considering that the previous Anchor-based method will bring a large number of parameters and the detection effect is not good.The recent recognition of the "Anchor-free" method by many scholars and the good experimental results.This paper constructs a bounding box regression algorithm based on "Anchor-free" in the head network part and designs a loss function to calculate the error between the predicted frame and the real frame.Thirdly,in view of the problem of inaccurate instance segmentation caused by target occlusion,target deformation,and multi-scale in different scenarios,this paper designs a mask branch based on the backbone network constructed before,parallel to target classification and bounding box regression Branch.To represent the error between the prediction mask and the real mask,we designed a mask loss function that accurately represents the error and has a small number of parameters.The mask branch's segmentation accuracy mainly depends on the ingenious design of the multi-path connection mode of the backbone network and the neck network and the appropriateness of the mask branch.In summary,this paper has carried out in-depth research from multiple angles such as model structure,sample equalization,loss function design,feature extraction,and fusion,and multi-path connection methods for target detection and instance segmentation,effectively improving the accuracy of object detection and the accuracy of instance segmentation.Experimental results show that the method proposed in this paper effectively improves the efficiency of target detection and instance segmentation,and has certain scientific research reference value and practical application value.