Research On Segmentation And Pose Measurement Algorithm Of Industrial Parts Based On Deep Learning

With the vigorous development of deep learning technology,more and more scenarios can use deep learning technology to achieve automated and intelligent operations,thereby reducing labor costs.The picking system,which is widely used in manufacturing and logistics industries,is at the forefront of using deep learning technology.The deep learning tasks mainly involved in the picking system include target detection,instance segmentation,and pose measurement.At present,major research institutions and major universities have proposed their own solutions for the picking system.However,in industrial scenarios,the accuracy of each algorithm needs to be improved for industrial parts that are disorderly placed,less textured,and heavily occluded..Therefore,we used the intelligent picking system in the industrial scene as the research background to study the pose measurement of industrial parts.Instance segmentation was determined to be a key factor,so a CenterMask-based instance segmentation network and pose measurement network based on DenseFusion were utilized,and related experiments were carried out to verify the practicability and accuracy of the improvement.First,we explored the basic knowledge related to deep learning,and focuses on related concepts in convolutional neural networks such as Convolutional layer,Pooling layer,Loss function and Optimization method,and analyzes the advantages and disadvantages of existing research.Then we introduced a depth information processing method which can retain the three-dimensional information of the scene and can use 2D convolution operations to extract data features.Next,we focused on a feature fusion algorithm based on the attention mechanism--the algorithm can efficiently fuse RGB features and depth features,simultaneously fusing features while suppressing features with low performance capabilities through the attention mechanism,clarifying the feature channels that the network model needs to pay attention to,enhancing the performance capabilities of features,and improving the performance of the segmentation network.Finally,we improved on the existing research results concerning the CenterMask instance segmentation network and DenseFusion pose measurement network.It introduces module by module the composition of the network used in the project and demonstrates through experiments that the improved network model can be used in industry.When the appropriate threshold is selected,the improved CenterMask network can segment more parts in a single frame image,and the segmentation speed is about 10 frames per second;measured by the improved DenseFusion,the ADD-S value is almost guaranteed to be below 1.8 mm.Applying the improved network model to the picking system,through multiple experiments,within 30 minutes,the robotic arm can perform an average of 122 grasping operations,and the grasping success rate is about 97%.The current research results still leave room for more in-depth research.