Multimodal Machine Learning Based Object Recognition Techniques for Engineering Applications

Object recognition technology has important applications in the industrial field, but there are limitations in relying on single modal information. To address this problem, the study proposes a multimodal object recognition system that combines tactile sensors and visual information, and processes tactile data through graph convolutional neural networks to improve recognition accuracy. The system utilizes multi-channel impulse map convolutional layer, attention layer and fully connected layer for tactile data analysis, and introduces Harris corner point detection algorithm, orientation gradient histogram and optical flow field orientation histogram to enhance visual feature extraction. The experimental results show that the recognition accuracy of the multimodal neural network system on different datasets reaches 95.85%, which is better than the performance of traditional KNN and DTW algorithms. The research system demonstrates high accuracy and stability in error analysis and physical localization experiments, for example, the error is only 0.1 cm in distance measurement experiments, and the attitude angle error is within 3 degrees. By combining haptic and visual information, the system realizes a more comprehensive and precise recognition of objects, which provides strong support for industrial automation and intelligent manufacturing. In summary, the effectiveness and advantages of multimodal fusion technology in object recognition tasks are remarkable and have a wide range of application potential.