Development And Application Of Atrous Difference Convolution Network Based Real-Time Semantic Segmentation Algorithm

At present,the semantic segmentation based on deep learning develops rapidly,and has broad application prospects in the fields of automatic driving,virtual reality and augmented reality.Most modern networks for semantic segmentation aim to improve accuracy,requiring a lot of floating point operations and long run-times.on the other hand,the networks mainly compromise spatial resolution to achieve real-time inference speed,which leads to degradation of accuracy.In order to push the real-time semantic segmentation network to higher precision and speed,this thesis develops the Atrous Convolution Difference Network(ACDNet),which includes Context Path and Spatial Path.The Context Path uses ResNet-18 as the backbone.The Atrous Difference Module(ADM)differentiates the local features from the environmental context features,further obtains global context information,and selects valid channels of feature map through the attention mechanism.The Spatial Path stacks several eonvolutions with small stride to eneode rieh spatial information.The Feature Selection Module(FSM)integrates the features of the two branches reasonably,which avoids the precision degradation caused by directly concatenated different types of features.ACDNet can achieve 67.7%mloU on the CamVid test dataset(960 × 720)at the speed of 81FPS on one NVIDIA GTX 1080Ti.For an input image with 1920 ×1080 resolution,forward inference of ACDNet only consumes 31.3ms.Based on ACDNet,this thesis develops a USB port defect detection algorithm,which runs on an embedded system with NVIDIA Tegra XI as the processor.First,ACDNet is adapted to improve the accuracy of small object,which achieves 95.9%mloU on the USB segmentation dataset.And this thesis made three improvements to Faster R-CNN,including adding anchors' type,adding multi-scale ROI Pooling,and using the feature pyramid network as the backbone.Then the improved Faster R-CNN is used to detect detection of the surface area,which can quickly and accurately catch detects such as pits,scratches and dirt.After analysis,the accuracy of the algorithm in this thesis is 96.23%,and the missed detection rate is only 1.29%.Therefore,the algorithm of this thesis has research significance and engineering application value.