| With the explosive growth of medical images to be detected,it is more and more difficult to detect whether there is tumor in the image quickly and accurately only by manual recognition.Therefore,image processing technology has developed rapidly.In medical image processing,the use of computers to quickly and accurately realize medical image detection and positioning can help experts more accurately control the condition of the disease.Among them,object detection technology is an important part.Object detection technology is the above-mentioned process of recognizing and detecting object behavior.Whenever a object(such as tumor)appears in the image,it defines a region and predicts its slave category to quickly and accurately implement the object.After the object is detected,the image to be detected will eventually become the detected image.Different from natural images,the edge of most tumor detection objects in medical detection images is often fuzzy and irregular,and the number of pixels is small.Such objects are small objects.For small objects,it is very difficult to accurately locate the object only by observing the image,which seriously affects the fast and accurate judgment.At present,object detection methods based on deep neural network have been widely used in medical images,but most of the detection tasks are based on specific scenes,which lack certain universality;in addition,most of the algorithms are not friendly to small objects.How to recognize small objects more accurately and make the algorithm more universal is a problem to be solved.In this paper,dense block structure is applied to the detection task.Combined with the aggregation theory,the image object detection is deeply studied,and the following two improved models are proposed.(1)A object detection model combining high / low dimensional dense features is constructed.Combined with the soft tissue structure and tumor characteristics in medical image sequence,the dense structure is applied to the tumor detection task,and the feature3 D mapping is established.In order to improve the robustness of the network and reduce the irreversible loss of initial features,feature 3D mapping technology is used to improve the network parameters,reuse image features and 3D image sequence,and then the tumor3 D sequence is transformed into hard threshold 3D to establish feature connection.Using aggregation method for feature reconstruction,fusion channel features,spatial features,fully mining the image pixel level feature association degree,making full use of detail features to avoid the loss of narrow features,improving the ability of network feature extraction and enhancing the ability of feature expression.In this paper,a variety of network structures and parameter combinations are designed,and experiments show that the network structure has achieved better results in the accuracy,average intersection ratio and average precision.(2)A multi-scale dense feature object detection model is constructed.In order to accurately and effectively distinguish different scale objects in natural images,in order to complete the correct recognition of features,inspired by small sample detection and multi-scale feature detection,a target detection network based on multi-scale feature is designed.Different from the traditional multi-scale feature extraction method,the network focuses on high-level feature detection and designs dense feature cascade switching domain.In addition,in order to obtain better accuracy,swish is added to the dense structure to improve the accuracy,deepen the abstract feature information,accelerate the network cycle,and reduce the computational consumption of requirements.In order to achieve excellent performance and computational efficiency,the effectiveness of the model under different networks and parameters is verified in experiments.Compared with the existing methods,this architecture achieves better results in the performance indexes such as the average intersection ratio and average precision. |
PDF Full Text Request