Parallel Algorithm Of Detection And Segmentation Based On Deep Learning

Object detection and semantic segmentation algorithms have always been one of the main research directions in the field of computer vision.Especially in the research of image parsing,it is necessary to consider these two tasks to work together at the same time,which often requires a high processor for the operation.In order to reduce the computational cost of the two tasks working together,this paper proposes a parallel algorithm for detection and segmentation.In recent years,the object detection algorithm and semantic segmentation algorithm based on deep learning have made great progress in both recognition accuracy and computational efficiency.There are many applications in industry.In this paper,parallel computing is proposed to make two subtasks work together.The design method uses deep learning techniques to propose improved design ideas for the algorithm itself.This paper first discusses the object detection and semantic segmentation algorithms based on deep learning in recent years and studies the current network structure design and working principle with excellent performance.The key point is to find out the common design ideas between the two algorithms.The similarities and differences between the algorithms in the decision-making layer are also made into the feasibility analysis of parallel algorithms.Then,according to the commonality of the two task structures,the parallel algorithm is divided into two aspects: feature extraction layer and decision-making layer.The feature of the detection and segmentation is shared by a feature extraction network to represent the feature information required by the two tasks.Thereby achieving the purpose of computing two tasks in parallel.Then,according to the difference of the two task structures,the decision layer is divided into three parts: a candidate bounding box classifier,a candidate bounding box regressor,and a segmentation mask generator,and according to each sub-task in the parallel algorithm,the utilization degree is expressed for the feature information.Dynamically adjust the complexity of the decision-making layer network map structure allocation,so as to meet the accuracy requirements of the model for the two tasks of detection and segmentation.Finally,using the transfer learning model training method,the classification network model parameters of the feature extraction layer are loaded into the network model of the parallel algorithm for model hyperparameter setting,training,and testing.In addition,this paper uses the innovative design method of deep learning technology in the convolutional neural network structure for the feature extraction layer,candidate frame regression,and segmentation mask generator.Using the hybrid codec design idea,the basic network of information fusion is proposed to characterize the feature information of the parallel algorithm.Using the decoder structure output candidate frame position parameters to correct the position parameters of the encoder two-class candidate frame,the spatial pyramid pooling method based on the atrous convolution and the deformation pooling alignment method are used to characterize the pixel information of the segmentation masker,and finally The feature extraction layer and the decision layer are combined by a connector to propose a parallel algorithm based on scale-invariant detection and segmentation.Through validation on standard datasets containing common annotated information in target detection and semantics segmentation tasks,this paper proposes a parallel algorithm for feature extraction based on information fusion and scale invariant detection and segmentation.Compared with the current mainstream target detection and semantics segmentation algorithms,on the premise of maintaining the accuracy of the current mainstream model,it improves the computing time and occupies the computing resources of processors.