1. Introduction
Direct laser deposition (DLD) is an advanced manufacturing technology that utilizes high-energy lasers to melt and deposit metal powders simultaneously onto three-dimensional parts [
1]. This technique offers advantages such as superior part performance, high manufacturing flexibility, short production cycles, and low costs, attracting significant attention from both academia and industry [
2]. However, as a complex process with numerous variables [
3], DLD faces challenges, such as poor process continuity and limited portability. Therefore, online measurements of the melt pool during deposition are essential for improving quality and yield.
The dimensions of the melt pool significantly affect the morphology of the deposited layer. During DLD, the melt pool exhibits high brightness, high temperature, small size, and rapid changes, with interference from sparks and plasma. Noncontact sensors based on machine vision offer a mainstream solution to these measurement challenges, allowing precise acquisition of essential melt-pool data.
Thermal and infrared imaging are the most common noncontact measurement methods. Criales et al. [
4] used a thermal camera to observe the geometrical features and splattering behavior of nickel alloy 625 during laser-powder bed fusion (L-PBF). Hu and Kovacevic employed [
5] a coaxial camera to capture infrared images of the melt pool, with the infrared signals fed back to a control system to regulate laser power and maintain a stable melt-pool width. Price et al. [
6] captured melt-pool images in electron beam additive manufacturing (EBAM) of Ti-6Al-4V powder using near-infrared (NIR) thermal imaging to analyze the temperature distribution and dimensions of the melt pool. Heigel and Lane [
7] identified liquid–solid transition discontinuities in an L-PBF melt pool by utilizing the local minima of the second derivative of the grayscale intensity signals captured by an IR high-speed camera. They then determined the overall shape of the melt pool by analyzing the corresponding intensity values. Cheng et al. [
8] collected radiation temperature data with an NIR thermal imager during selective laser melting and utilized the collected thermal images and radiation temperature curves to calculate the liquid–solid phase transition lines and extract melt-pool width information. Da Silva et al. [
9] used a laser coaxial thermal imager to detect overhead melt-pool contours and extract geometric dimensions, such as length and width. However, the high cost of high-speed, high-resolution NIR cameras limits their practical application.
With advancements in semiconductor devices and integrated circuit technology, relatively low-cost image sensors, such as charge-coupled devices (CCDs) and complementary metal-oxide semiconductors (CMOSs) have become widely used. Machine-vision-based melt-pool image measurements primarily focus on dimensions, such as the width and height of the melt pool. Kim and Ahn [
10] developed a monitoring system with a coaxial illumination laser and a CCD camera equipped with optical filters designed to monitor the two-dimensional keyhole shape in Yb: YAG laser welding. Doubenskaia et al. [
11] demonstrated that in the Ti6Al4V laser cladding process, grayscale image brightness in the liquid–solid transition zones is consistent across different cladding parameters, providing a reliable method for estimating temperature distribution and calculating dimensions of the melt pool. Yang et al. [
12] developed an online system using a side-positioned CCD camera with neutral-density filters and image-processing algorithms to eliminate noise and accurately determine melt-pool width. Hao et al. [
13] introduced a weld pool imaging system based on spatial filtering and the Abbe imaging theory, using a high-pass spatial filter to eliminate low-frequency background noise and precisely capture the contours of the melt pool, weld seam, and arc. Le et al. [
14] established a machine-vision-based laser selective melting system for online melt-pool dimension measurement with a lateral off-axis setup. Devesse et al. [
15] developed a field programmable gate array-based image capture system that transferred sensor data to a computer for processing with MATLAB simulations to predict melt-pool dimensions. However, these single-camera side-shot methods are limited to detecting melt pools in one deposition direction.
Asselin et al. [
16] used a trinocular vision system with cameras set 120° apart to monitor the melt pool from various angles, applying perspective transformation algorithms to measure melt-pool widths from different directions; however, they overlooked edge position changes due to varying deposition directions. Mazzoleni et al. [
17] proposed a coaxial imaging system with an external light source monitor module, CMOS camera, and filters. Yang [
18] designed a coaxial camera system that measured melt-pool dimensions by capturing the entire upper surface of the melt pool with a coaxially aligned camera and using the minimum bounding rectangle method for dimension extraction.
Recently, machine-learning-based neural network techniques have been applied to the measurement of melt-pool images. Yu et al. [
19] utilized an enhanced U-Net architecture network trained with limited sample data for real-time recognition and tracking of arc-welding melt-pool boundaries. Wang et al. [
20] input melt-pool images and simultaneously collected weld voltage signals transformed via short-time Fourier transform (STFT) into a CNN to identify welding defects. However, these neural networks only achieved target detection and were significantly affected by noise, reducing accuracy in melt-pool dimension measurements. Jiang et al. [
21] developed a backpropagation neural network prediction model for laser cladding height using process parameters (laser power, powder feed rate, and scanning speed) to predict melt-pool height. Building on the work by Asselin et al. [
16], Ravani-Tabrizipour et al. [
22] input information, such as the angles of the major and minor axes of the melt-pool ellipse, obtained from multicamera projection transformations, into an algorithm combining image tracking protocols and recursive neural networks to detect the overlay height of SS303L deposited on low-carbon steel under various conditions. However, the output represents the height of the deposited overlay, which differs from that of the liquid melt pool during deposition. She et al. [
23] built a multisensor monitoring system for the laser welding process, extracted the image features of the molten pool as well as laser-induced plasma spectral features, and established a real-time prediction model for the high-precision welding depth of fusion.
Most machine-vision-based methods for detecting melt-pool width [
12,
13,
14,
15] rely on off-axis side-mounted monocular cameras, which are practical only for melt pools deposited along the optical axis of the camera, limiting their utility. When the deposition direction of the melt pool is not parallel to the camera setup, the limitations of the camera’s projection detection principle result in detecting only the near edge while the far edge is obscured, capturing a “false edge” (
Figure 1) projection on the upper surface of the melt pool.
To improve the measurement range and accuracy of melt-pool width and prevent incorrect width extraction due to false edges, this study proposes a binocular-vision-based method employing off-axis cameras. Binocular cameras detect both edges of the melt pool, and the images are aligned in the same spatial plane coordinate system using perspective and affine transformations. By incorporating deposition angle data, this method achieves accurate melt-pool width measurements.