Research On Object Recognition Algorithm For Automatic Assembly Of Electronic Components

With the rise of computer vision,object recognition technology is widely used in industrial production,but the problem of small object recognition and algorithm accuracy and speed in object recognition applications remains to be solved.Relying on the National Natural Science Foundation project "Intelligent Assembly Robot Visual Autonomous Recognition,High-precision Positioning and Compliant Control Method Research",this paper focuses on the object recognition algorithm for automatic assembly of electronic components,by analyzing recognition algorithms in electronic component recognition,improve the original SSD algorithm,and finally realize high-precision real-time identification and detection of electronic components in complex scenes,which lays a foundation for the assembly of electronic components.The main research contents of this paper are as follows:?1?The basic theory of target recognition in convolutional neural networks is discussed,and the existing target recognition algorithms and their difficult problems in electronic component identification are analyzed.Firstly,the convolutional neural network model is introduced,as well as the basic theories such as regional suggestion,frame regression and non-maximum suppression in target recognition.Then,the effects of each recognition algorithm in the recognition of electronic components are compared by experiments.Good SSD is the basic algorithm of this paper.At the same time,it discusses two difficult problems of small target recognition and algorithm precision and speed in electronic component recognition.?2?Based on the original SSD algorithm,an improved SSD algorithm based on feature separation and target context?called CO-SSD?is proposed for the identification of small targets in electronic component recognition.Firstly,the feature map network for classification prediction and location prediction in the SSD network is separated to realize the separation of the target location task and the classification task.Secondly,the feature maps of the fc7,conv8₂ and conv9₂ layers in the classification feature map are performed on the basis of the two.Double upsampling,adding the context information of the target to make up for the insufficient feature information of the small target;finally,multi-scale feature fusion is performed to output the prediction result.Experiments show that compared with the original SSD algorithm,the mAP value of the CO-SSD algorithm is increased by 1.1%-1.3%.?3?It is difficult to balance the recognition accuracy and speed in the identification of electronic components.Based on the CO-SSD algorithm,an improved CO-SSD algorithm based on feature enhancement?called CCO-SSD?is proposed.The selective enhancement module is added to the classification feature map network of the original CO-SSD algorithm to enhance the effective feature channel in the feature layer and suppress the invalid feature channel in the feature layer,thereby further improving the recognition accuracy and reducing redundant information;The mAP value of CCO-SSD is further improved than the CO-SSD algorithm,and the algorithm speed basically reaches the speed of the SSD algorithm,achieving both recognition accuracy and speed.