Research On Multi-scale Object Detection And Segmentation Algorithms Based On Deep Learning

Deep learning technology has benefited from the improvement of computing capability and the big data era coming,has made major breakthroughs.Especially the field of computer vision based on convolutional neural networks has developed rapidly,and has achieved much better results than traditional algorithms on multiple tasks.However,because deep learning algorithms require massive amounts of data for training,the multi-scale objects bring new challenges.Usually,researchers solve the multi-scale problem with three aspects: convolutional network architecture design,data processing,and convolution operations.In the terms of convolutional network architecture design,parallel design allows the network to extract the multi-scale features of the object while using as few parameters as possible,serial design allows the fusion of feature information at different scales,and multiscale prediction is performed on feature maps at different scales prediction.Based on this,this paper designs a lightweight multi-scale object detection network structure,it considers the computational complexity and result,and obtains a win-win result on the COCO dataset.In terms of data processing,algorithms based on cropping and scaling enable multi-scale objects to be predicted at the same scale.In the face of extremely small objects and extremely large images,through effective slicing operations,small targets can be predicted directly without downsampling.Based on the slicing algorithm,this paper proposes a slicing fusion algorithm,which can merge the segmented objects into a complete object,and solves the problem that the original slicing algorithm can not effectively detect the multi-scale object in the large image.In terms of convolution operations,deformable convolution effectively solves the problem of fixed sampling points of a single convolution layer by adding an offset to the sampling points of the convolution kernel,but it also brings the disadvantage of complicated use and retraining.This paper analyzes the time cost and effect of deformable convolution operation and standard convolution operation under different factors,and proposes an algorithm that uses deformable convolution to upgrade the original network,which can be easily upgraded to any standard using deformable convolution Convolutional network,and on the PASCAL VOC data set and COCO data set get a better result.