Abstract: For the problems about low accuracy and poor robustness under complex conditions such as adhesion, stacking, illumination change and environmental factors, a part real-time detection method based on improved YOLOv3 deep learning framework is proposed. The CFE module is embedded in each residual network of Darknet-53 basic feature extraction network. A deep feature extraction network CFE-Darknet53 that combines CFE module and Darknet-53 is constructed. A real-time part detection model based on CFE-Darknet53 under the YOLOv3 deep learning framework is established, which improves the feature extraction ability of the detection network in complex environment. An improved K-means algorithm is designed to predict the bounding box. The data sets of parts are clustered to select the optimal number and size of anchor frames, which further improves the detection accuracy. The experimental results show that the detection accuracy of the improved YOLOv3 algorithm for multi class parts with high similarity can reach 91.6% under complex conditions. Compared with the YOLOv3 algorithm, the improved YOLOv3 algorithm improves the detection accuracy by more than 10%. The detection time is 43 ms, so the improved method achieves real-time detection of parts at the video transmission frame rate of 24 frames/s.

Key words: robot, deeplearning, real-timedetection, featureextraction, clusteranalysis