Research on Robot Target Classification and Localization Based on Improved Mask R‐CNN

The small workpieces are easily missed during detection, and the irregular workpieces are difficult to recognize and segment effectively by traditional detection algorithms in the industrial field. The traditional target detection algorithms have problems such as low accuracy and poor generalization performance. This paper proposes a robot target recognition and positioning method based on the improved Mask R‐CNN. First, the network structure is designed to add a Convolutional Block Attention Module (CBAM) in the backbone, replace the Feature Pyramid Network (FPN) structure used in the original model of Mask R‐CNN with a Path Aggregation Network (PAN) structure, and increase the receptive field to enhance the recognition of small target objects and the segmentation of multi‐objects. Second, after classification is completed, according to the segmentation information, the output is augmented with center coordinates and rotation angle information. Finally, comparative experiments are conducted in the COCO dataset and the industrial part dataset to verify the effectiveness and practicality of the proposed algorithm. The experimental results show that the improved model achieves an AP 50 of 60.6 in the COCO dataset and 99.4 in the industrial parts dataset. Additionally, in single‐object and multi‐object grasping experiments, the grasping accuracy is 91.5% and 85.3%, respectively.