Research On Rru Power Port Detection Technology Based On Deep Neural Network

The vision robot captures the image of the surrounding environment through the camera,obtains the position and orientation information of the object based on the im-age,and the manipulator is controlled to complete the corresponding operation.It is very important to get the position and orientation information of the object accurately and quickly through the image.Specifically in the application of this paper,the goal is to obtain the position and orientation information of the power port on the RRU product.Considering that traditional object detection techniques need to extract features manu-ally,the accuracy and efficiency in complex background are relatively low,this paper chooses a deep neural network with good robustness and high accuracy to detect the object.In this paper,a method combining neural network with traditional image pro-cessing technology is proposed to obtain the position and orientation information of power port.Firstly,the bounding boxes of two pins in the power port are predicted by using the neural network.Then,the image regions of predicted bounding boxes are bi-narized by the Otsu algorithm,and the image moments are used to get the two pin cen-ters,which indirectly obtains the position and orientation information of the power port.Considering that camera distortion will cause pin shape distortion in the image,which will lead to errors in obtaining the center point based on the pixel information in the image regions of predicted bounding boxes.In this paper,the distortion model for the camera is established to correct the image.Because the object to be detected is a pin,there is no universal dataset,so it is nec-essary to establish the corresponding dataset.In this paper,the specific process of da-taset establishment and the means of data augmentation used in the network training process are introduced.Data augmentation means include random cropping,clipping,deflection,scaling,color jittering and so on.The neural network model for object detection is improved based on CornerNet.The specific improvements include the following four aspects.Considering the com-plexity of CornerNet's feature extraction network structure,there are a lot of computing costs.In this paper DenseNet with dense connection is used as the basic feature extrac-tion network to realize feature reuse and enhancement.The standard convolution in the network is replaced by the deep separable convolution,which reduces the network pa-rameters and computational complexity.Squeeze-and-Excitation blocks are introduced to enhance channel-wise feature responses.Continuous deformable convolution layers are introduced to make the convolution receptive field adjust adaptively with the size and shape of the object to be detected.In addition,the selection of network optimization method and the improvement of non-maximum suppression algorithm are discussed.Optimizer with dynamic learning rate boundary is more effective in parameter updating,and by using the improved soft non-maximum suppression algorithm to remove redun-dant bounding boxes the detection results can be effectively improved.The experimental results show that using the same parameter optimization method and non-maximum suppression algorithm,the number of the improved optimal network parameters is greatly reduced,which is about 1/28 times of the original network.AP of the improved optimal neural network model is improved by 7.2%compared with the original network model,AP⁵⁰ andAP⁷⁵ reach 95.8%and 92.7%respectively based on the test dataset.