Dahang Wan

Fringe projection technology is a commonly used technique in optical three-dimensional measurement. In high-speed motion scenarios, due to image noise and the effects of object motion, projecting more fringe patterns for high-precision phase unwrapping is a common method, which can significantly reduce the frame rate of three-dimensional reconstruction. Deep learning techniques have been employed for high-precision phase unwrapping, but typically these models have a large parameter and computation, making them difficult to integrate into real-time three-dimensional reconstruction systems. In this paper, we first employ the LUT technique for rapid computation of dual-frequency phases. Secondly, we design a deep learning model with a parameter size of only 276kb for high-precision phase unwrapping, and quickly embed it into a real-time three-dimensional reconstruction system through 8bit quantization without compromising accuracy. Furthermore, we utilize the calibration parameters of a real FPP system to establish a corresponding virtual FPP system for rapid generation of data required for model training. Lastly, we optimize the generation of point clouds by avoiding the computationally slow inverse matrix operation process. Experiments show that our model can achieve high-precision real-time three-dimensional reconstruction at a rate of 130 frames per second.

Object detection is essential to the interpretation of optical remote sensing images and can serve as a foundation for research into additional visual tasks that utilize remote sensing. However, the object detection network currently employed in optical remote sensing images underutilizes the output of the feature pyramid, so there remains potential for an improved detection. At present, a suitable balance between the detection efficiency and detection effect is difficult to attain. This paper proposes an enhanced YOLOv5 algorithm for object detection in high-resolution optical remote sensing images, utilizing multiple layers of the feature pyramid, a multi-detection-head strategy, and a hybrid attention module to improve the effect of object-detection networks for use with optical remote sensing images. According to the SIMD dataset, the mAP of the proposed method was 2.2% better than YOLOv5 and 8.48% better than YOLOX, achieving an improved balance between the detection effect and speed.

Magnetic rings are widely used in automotive, home appliances, and consumer electronics. Due to the materials used, processing techniques, and other factors, there will be top cracks, internal cracks, adhesion, and other defects on individual magnetic rings during the manufacturing process. To find such defects, the most sophisticated YOLOv5 target identification algorithm is frequently utilized. However, it has problems such as high computation, sluggish detection, and a large model size. This work suggests an enhanced lightweight YOLOv5 (MR-YOLO) approach for the identification of magnetic ring surface defects to address these issues. To decrease the floating-point operation (FLOP) in the feature channel fusion process and enhance the performance of feature expression, the YOLOv5 neck network was added to the Mobilenetv3 module. To improve the robustness of the algorithm, a Mosaic data enhancement technique was applied. Moreover, in order to increase the network’s interest in minor defects, the SE attention module is inserted into the backbone network to replace the SPPF module with substantially more calculations. Finally, to further increase the new network’s accuracy and training speed, we substituted the original CIoU-Ioss for SIoU-Loss. According to the test, the FLOP and Params of the modified network model decreased by 59.4% and 47.9%, respectively; the reasoning speed increased by 16.6%, the model’s size decreased by 48.1%, and the mAP only lost by 0.3%. The effectiveness and superiority of this method are proved by an analysis and comparison of examples.

During the detection phase of computer vision in industry, insufficient data occurs frequently. In this scenario, data augmentation can effectively resolve the issue and improve the performance of the model. Presently, image data augmentation has become a standard model configuration, but few image data augmentation approaches are available for industrial object detection. This paper introduces a new method for augmenting picture data, Random Interpolation Resize (RIR), which can enhance the detection effect of industrial data sets without the requirement for extra resources. The resize operation is a standard configuration in the preprocessing stage, so we alter the interpolation method from fixed collocation to randomly selected interpolation to increase the amount of image data, enhance the model's generalizability, and enhance its detecting capabilities. Multiple data sets were utilised in experiments. The mAP of the YOLOv5s model on the GC 10 dataset increased by 4.8%, the mAP of the Yolov5n model on the Neo-DET dataset increased by 1.4%, and the mAP of the YOLOv5s model on the battery steel shell dataset climbed by 0.9%. The mAP of the YOLOv5 model was increased by 0.5% when applied to the VOC dataset.