Laser Spot Edge Extraction and Positioning Optimization Based on Multi-Scale Adaptive Convolution

Abstract

Computer vision and image processing Laser spot localization is an important application of computer vision in industrial inspection applications, automated control, and other applications in the fields. The conventional spot localization based on the fixed-scale convolutional networks is usually an edge detector and localizer, but it cannot perform well on detecting and localizing complex, overlapping, or incomplete spots. In order to address these drawbacks, we will introduce a new laser spot localization technique using multi-scale adaptive convolution module and edge extraction and localization optimization module (MSACM + EELOM). The most significant value of this work is the introduction of a multi-scale adaptive convolution module dynamically changing the size of the kernel to meet the fluctuations of the spot scale and an edge localization optimization module that improves the alignment of the boundaries to minimize the errors of the noise and background complexity. Moreover, a new loss is developed that combines edge consistency loss and localization accuracy loss to balance between accuracy in extraction of edges and localization of spots. Ablation experiments prove that each of the modules is working, and the results obtained in the course of the experiments in general ensure that our whole model can localize with 97.96 percent accuracy, falling far ahead of the traditional methods in terms of accuracy, robustness, and adaptation.