Research And Application Of Image Object Detection Algorithm Based On Deep Learning

With the advent of the big data era,more and more data needs to be stored and processed.This not only requires massive storage devices,but also requires computers with powerful data processing capabilities to complete data processing tasks efficiently.In this era,deep learning came into being and quickly applied to various fields such as manufacturing,medicine,transportation,and finance.The rapid development of deep learning technology and the widespread use of large-scale annotated image datasets have made computer vision technology a significant breakthrough in image processing,especially in the field of object detection.The task of object detection in computer vision mainly includes two parts:the positioning prediction and classification of target objects.Object positioning is to extract features of the target object,and locate the extracted features with the anchor-based or anchor-free model,and then predict the position of the target object in the image.After that,the object classification module classifies the detected objects and obtains the final classification result.On the one hand,in the two-dimensional object detection task,the image composition is generally more complicated(for example,the image contains a single or many target objects,the image background information is complex,and the target objects are difficult to distinguish),and the object detection task will become relatively difficult.On the other hand,from the perspective of a person,the objects in the image contain not only two-dimensional information,but also corresponding three-dimensional information.Therefore,while studying the extraction of two-dimensional object information,it is also necessary to extract the three-dimensional information of the target object to reconstruct the three-dimensional target object.Based on this concept,this paper proposes an improved object detection algorithm based on deep learning for two-dimensional object detection tasks,joint 2D object detection and 3D reconstruction tasks,respectively.For the task of 2D object detection,this paper proposes an object detection algorithm based on a new cascade parallel detector Grid R-CNN.The algorithm constructs detectors by paralleling anchor-based branches and anchor-free branches.In the anchor-based branch,the detector combines the object location method of Grid R-CNN with the shape prediction module method of GA-RPN to predict the more accurate anchor box shape.In the anchor-free branch,the detector reduces the number of overlapping anchor boxes by paralleling the FSAF branch,thereby obtaining high-quality anchor boxes.Finally,this paper uses the detector cascade method to complete the anchor box determination and object classification by training a detector that continuously increases the Intersection-over-Union(IoU)threshold,thereby further improving the accuracy of object detection.The experimental results show that the performance of our proposed algorithm on the Pascal VOC2007 and COCO2017 is generally better than the comparison algorithm.In terms of joint 2D object detection and 3D reconstruction tasks,this paper proposes a novel joint 2D object detection and 3D reconstruction algorithm based on the adversarial fusion Mesh R-CNN.The algorithm first uses the DCGAN module to generate adversarial images based on real images,and then uses GA-RPN for 2D object detection.To obtain robust voxels,the algorithm uses Pix2Vox to convert 2D pixels to 3D voxels and perform voxel fusion to improve the quality of 3D voxels generation.Finally,we use vertex alignment and the Principal Neighborhood Aggregation network(PNA)to refine the 3D mesh and obtain the final 3D mesh model of the target object.The experimental results show that the performance of our proposed algorithm on the Pix3D dataset is generally better than the comparison algorithm.This paper proposes improved algorithms for the 2D object detection task,the joint 2D object detection and 3D reconstruction tasks,respectively.The algorithms have certain theoretical innovation and application value,and have obtained better detection results in experiments.