Research Of Deep Learning Based On Upsampling Technology

At present,most of the upsampling operators based on deep learning are prone to generate artifacts and blurs in regions with rich texture features when magnifying the size of the feature map which cannot distinguish the importance of extracting the feature information,resulting in the inability to effectively extract the details contained in the feature maps and influencing the effect of the prediction tasks.In view of the above problems,this paper discusses the design ideas of effective upsampling and proposes a feature-enhanced upsampling method,constructing different model structures in semantic segmentation and object detection networks.Specific work is as follows :Firstly,in order to avoid the blurs of segmentation images caused by simple upsampling operators,this paper proposes a feature-enhanced upsampling operator to generate a region-adaptive upsampling kernel,which can filter effective information in dense channels and eliminate the interference of redundant information.The feature-enhanced upsampling kernel reassembles the features in a content-aware manner,and reconstructs the output with clear texture while restoring the resolution of the feature maps.In addition,starting with the feature pyramid network,this paper proposes a multi-scale fusion method of jump structure,which supplements high-level semantic information to low-level,making up the difference of information between different levels,and strengthening the semantic information expression of small objects.Experiments on classical semantic segmentation tasks show that the proposed algorithm has considerable effect on ADE20 k scene dataset,which can be used as a strong basis for future research.Then,in order to improve the detection effect of object detection algorithms on multi-scale object in images,this paper disscusses the design of feature-enhanced upsampling operator and designs an optimized feature-enhanced upsampling operator.The optimized feature-enhanced upsampling introduces the convolution block attention mechanism with channel and spatial perception ability to fully express the feature details,and uses the convolution operation with multi-scale receptive field to improve the multi-level local transformation ability of the kernel,strengthening the pixel recovery effect of different size features in the upsampling process,so as to improve the performance of multi-scale object detection.Besides,this paper proposes an improved feature pyramid network,combining horizontal with across-up paths to construct horizontal connections and introduces the parallel dimension reduction architecture to capture different receptive fields.By constructing fine resolution networks with cross-space dimensions,the performance of feature fusion networks is improved.Experiment data on PASCAL VOC dataset shows that the proposed algorithms have achieved good results on the object detection network.