CNN-Based Model for Pose Detection of Industrial PCB

For applications in robot manipulate with object, get the pose of objects is very important for controller's subsequent operations, especially in PCB feeding and blanking field, the grasp success rate will be enhanced if robot can get a exact pose of objects that relative to end manipulator. So in this paper we utilize the CNN model to build on a neural network for 3 tasks: object recognition, location and pose detection. This model treat pose detection as a classification problem and try to combine recognition, location at the same level. To validate the performance of the multi-task detection model, experiments and analysis of the model performance was carried out by the real-time PCB detection test. In the experiment, we use the PCB dataset comprised of 3 types which contains different poses made by ourselves as train/test samples. The number of object pose categories was divided into 8bins, 12bins and 36bins according to pose detection precision. We analysis the effect of the non-uniform datasets on training process and the final detect results shows that this CNN-based detection model can achieve high accuracy on PCB pose detection.