Yi Wang

For the small-size defects of metal workpiece surface and the difficult segmentation of image defects due to non-uniform illumination, an improved U-net semantic segmentation network was proposed to achieve accurate image segmentation of surface defects for metal workpieces. Firstly, the convolutional block attention module (CBAM) was integrated into the U-net netwok to improve the significance of the defective targets in the image. Secondly, part of the traditional convolution in the network was replaced by depthwise over-parameterized convolution (DO-Conv) to increase the number of learnable parameters of the network. Then, the Leaky Relu function was used instead of the partial Relu function in the network to improve the feature extraction ability of the model for the negative intervals. Finally, the median filtering and non-uniform illumination compensation method were used for image preprocessing, so as to reduce the effect of non-uniform illumination on the surface defects of metal workpiece images. The results show that the improved network mean intersection over union, accuracy rate and Dice coefficient index reaches 0.833 5, 0.933 2 and 0.867 4, respectively. The improved network significantly improves the segmentation effect of surface defect images of metal workpieces.

Aiming at the problem that Simultaneous localization and mapping (SLAM) is greatly disturbed by many dynamic elements in the actual environment, this paper proposes a real-time Visual SLAM (VSLAM) algorithm to deal with a dynamic indoor environment. Firstly, a lightweight YoloFastestV2 deep learning model combined with NCNN and Mobile Neural Network (MNN) inference frameworks is used to obtain preliminary semantic information of images. The dynamic feature points are removed according to epipolar constraint and dynamic properties of objects between consecutive frames. Since reducing the number of feature points after rejection affects the pose estimation, this paper innovatively combines Cylinder and Plane Extraction (CAPE) planar detection. We generate planes from depth maps and then introduce planar and in-plane point constraints into the nonlinear optimization of SLAM. Finally, the algorithm is tested on the publicly available TUM (RGB-D) dataset, and the average improvement in localization accuracy over ORB-SLAM2, DS-SLAM, and RDMO-SLAM is about 91.95%, 27.21%, and 30.30% under dynamic sequences, respectively. The single-frame tracking time of the whole system is only 42.68 ms, which is 44.1%, being 14.6–34.33% higher than DS-SLAM, RDMO-SLAM, and RDS-SLAM respectively. The system that we proposed significantly increases processing speed, performs better in real-time, and is easily deployed on various platforms.