Laser-based machine vision and volume segmentation algorithm for equal weight portioning of poultry and fish

Installation of the machine vision system

The machine vision system for this study was designed and constructed using an aluminum profile frame (510 mm × 560 mm × 730 mm). This structure supported the installation of the laser line profile sensor and the base for the sliding mechanism, which moved the object during the scanning process, as shown in Fig. 4A. The sliding mechanism provided a scanning range exceeding 150 mm, starting from the encoder installed rail end. The sensor’s laser was aligned at the upper boundary to optimize data capture. The sensor was positioned at 320 mm, producing a 172 mm scanning area, which was later adjusted to a target region of 150 × 150 mm. During scanning, the object moved upward along the sliding rail to ensure accurate data acquisition.

• (a)

  Aluminum frame and sliding object base for scanning. The red-highlighted area indicates the designated region for placing the object during scanning. The yellow arrow shows the scanning direction along the vertical axis as the object moves through the laser line profile sensor’s field of view.

• (b)

  System wiring and device connections including laser line profile sensor, encoder, control module, and computer interface used to acquire point cloud data.

• (c)

  Block diagram of the machine vision system. Encoder-generated triggers initiate laser profile capture, ensuring uniform spacing. The control module synchronizes sensor and encoder data before transmission to the personal computer (PC) for 3D reconstruction and cutting line computation.

To connect the laser line profile sensor to a computer and an encoder, as well as to supply power, a control module (Master 100 (LMI Technologies Inc, 2019)) was used. Figure 4B illustrates the interconnection of the sensor, encoder, control module, and computer. The I/O cordset linked the sensor’s I/O and Power/LAN ports to the control module, enabling signal transmission to the encoder. The power and Ethernet cordset connected the sensor to both the power supply and the computer, while the encoder’s signal cable was linked to the control module. Acting as a central hub, the control module supplied power and transmitted sensor data to the computer via Ethernet. The vision processing software, utilizing the Generic Interface for Cameras (GenICam) protocol, processed the data collected from the sensor, ensuring seamless hardware software communication.

To clarify the data communication and processing workflow, Fig. 4C presents a block diagram of the machine vision system. The process begins with the laser line profile sensor, which projects a laser line onto the surface of the object placed on a jig. A trigger event is required for the sensor to capture each profile image. In this setup, the trigger is generated by an incremental encoder, which detects motion as the object moves along the scanning axis. Once triggered, the laser is strobed, and the sensor camera captures a single image. This image is internally processed by the sensor to yield a profile (range/distance information). These profiles are sequentially collected to reconstruct a full 3D point cloud of the object. The control signal module receives inputs from both the sensor and the encoder and acts as a synchronization hub, managing data flow and power supply. Finally, the data is transmitted to the PC, where image processing software executes the cutting algorithm to calculate cutting lines.

The total processing time from initiating the scan to generating cutting lines, when dividing the object into 2 to 6 segments, was approximately 20 s. This duration was divided into three stages: first stage 10 s for initializing the system and establishing the connection between the laser line profile sensor and image processing software upon pressing the “Run” command, second stage 5 s for moving the object through the scanning range and rendering the point cloud, and third stage approximately 2–3 s for computing the cutting line positions using image processing software, excluding additional time for visual verification and display.

The laptop used in this research had the following specifications:

Processor: AMD Ryzen 5 5500U

Graphics Card: Radeon Graphics 2.10 GHz

Random Access Memory (RAM): 16 GB DDR4 3,200 MHz

Storage: SSD M.2 512 GB

Operating system: Windows 11

Volume computation

The point cloud data was collected using the laser line profile sensor, which captured point cloud data from the meat’s surface and transmitted the information as images to vision processing software. Each pixel in the captured image contained 3D coordinate data, which was extracted through programmed processing to reconstruct the full point cloud representation of the object (MVTec, 2022), as illustrated in Fig. 5. However, due to occlusions, surface reflections, or scanning limitations, certain regions of the point cloud contained missing data. To resolve this issue, the built-in Gap Filling function of the laser line profile sensor was employed. This gap filling method reconstructs missing data points by either selecting the lowest values from nearest neighbors or performing linear interpolation between neighboring points. Gap filling was applied along both the X- and Y-axes using a window size of x = 8 mm and y = 5 mm to ensure surface completeness. In addition, the built-in smoothing filters of the laser line profile sensor were used to reduce noise and enhance surface continuity. Smoothing operates by replacing each data point with the average of that point and its nearest neighbors within a defined window. In this study, smoothing was applied in both directions, with the X axis processed first followed by the Y axis, using a smoothing window size of x = 2 mm and y = 2 mm. This preprocessing step resulted in a cleaner, more consistent point cloud for subsequent volume computation and segmentation (LMI Technologies Inc, 2019).

Once the point cloud was obtained, the system applied Delaunay triangulation (Cheng, Dey & Shewchuk, 2012), creating a triangular surface mesh to model the 3D shape of the meat. This method ensured that each triangle was formed by connecting adjacent points, allowing for an accurate representation of the surface topology as illustrated in Fig. 6. The total volume of the object was then computed by summing the volumes of vertical triangular prisms formed between each triangle in the surface mesh and the reference base plane as illustrated in Fig. 7. For each triangle $\mathrm{k}$, the volume was calculated using Eq. (1).

(1) $${\mathrm{V}}_{\mathrm{k}}={\mathrm{A}}_{\mathrm{k}}\times {\mathrm{h}}_{\mathrm{k}},\mathrm{k}=1,2,3,\dots ,\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{t}\mathrm{t}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{n}\mathrm{g}\mathrm{l}\mathrm{e}\left(\mathrm{t}\right)$$where

${\mathrm{A}}_{\mathrm{k}}$ = the area of each triangle $\mathrm{k}$ on the surface mesh,

${\mathrm{h}}_{\mathrm{k}}$ = the average vertical distance (height) from the triangle’s face to the base plane.

The total volume of the object is then approximated by summing the volumes of all such prisms, as shown in Eq. (2).

(2) $${\mathrm{V}}_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}=\sum _{\mathrm{k}=1}^{\mathrm{t}}{\mathrm{A}}_{\mathrm{k}}\times {\mathrm{h}}_{\mathrm{k}},\mathrm{k}=1,2,3,\dots ,\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{t}\mathrm{t}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{n}\mathrm{g}\mathrm{l}\mathrm{e}\left(\mathrm{t}\right)$$

This total volume was later used in the cutting line determination process to ensure equal weight portions. Once the surface was modeled, the total meat volume was estimated by integrating the surface mesh over a defined reference plane.

Cutting line determination

After calculating the object’s volume, a volume-based method was applied to determine the cutting lines. These lines were defined in two ways: either as vertical cuts only or as a combination of vertical and horizontal cuts, ensuring that all segments had equal weight and volume. The position of vertical cutting lines were represented by the variable ${\mathrm{x}}_{\mathrm{i}}$, while the position of horizontal cutting lines were represented by ${\mathrm{y}}_{\mathrm{i}}.$ For example from Fig. 8, in the case where the object was divided using vertical cuts, the algorithm determined ${\mathrm{x}}_{\mathrm{i}}$and ${\mathrm{x}}_{\mathrm{i}+1}$ so that the volume of each segment ( $\mathrm{V}\left[{\mathrm{x}}_{\mathrm{i}},{\mathrm{x}}_{\mathrm{i}+1}\right]$ for $\mathrm{i}=0,1,2,\dots ,{\mathrm{n}}_{\mathrm{t}}$) remained equal.

The step-by-step process is explained through the pseudo code provided below.

Initialization (Lines 01–06): The algorithm defines the number of segments along the vertical ( ${\mathrm{n}}_{\mathrm{v}}$) and horizontal ( ${\mathrm{n}}_{\mathrm{h}}$) directions, calculates the total number of final segments ( ${\mathrm{n}}_{\mathrm{t}}$), and determines the target volume for each segment ( ${\mathrm{V}}_{\mathrm{I}}$). For each vertical segment, a target volume ( ${\mathrm{V}}_{\mathrm{i}}$) is computed based on its corresponding horizontal divisions.

For example, if ${\mathrm{n}}_{\mathrm{v}}$ = 3 and ${\mathrm{n}}_{\mathrm{h}}$ = [2, 3, 1], this configuration divides the object into three vertical regions. The first region is subdivided into two horizontal slices, the second into 3, and the third into 1, resulting in a total of ${\mathrm{n}}_{\mathrm{t}}$ = 6 segments. The total object volume is divided equally across these six segments, such that each piece should have volume ${\mathrm{V}}_{\mathrm{I}}$ = Total Volume/6. Consequently, the volume allocated for each vertical segment becomes ${\mathrm{V}}_{\mathrm{i}}=\left[2,3,1\right]\times {\mathrm{V}}_{\mathrm{I}}$.

Vertical segmentation (Lines 08–34, Fig. 9): In this step, the goal is to divide the 3D object into several vertical parts along the x-axis, each having a specific volume depending on ${\mathrm{V}}_{\mathrm{i}}$. The algorithm first identifies the object’s minimum and maximum x-coordinates ( ${\mathrm{x}}_{\mathrm{m}\mathrm{i}\mathrm{n}}$, ${\mathrm{x}}_{\mathrm{m}\mathrm{a}\mathrm{x}}$), indicated by the dashed lines in Fig. 9, left. It then searches for a cutting position ${\mathrm{x}}_{\mathrm{c}\mathrm{u}\mathrm{t}}$ (shown as the red line in Fig. 9, left) such that the volume between ${\mathrm{x}}_{\mathrm{m}\mathrm{i}\mathrm{n}}$ and ${\mathrm{x}}_{\mathrm{c}\mathrm{u}\mathrm{t}}$ closely matches the target (shown as the green block in Fig. 9, right). If the volume is too large, it adjusts to a smaller position ( $\mathrm{S}\mathrm{e}\mathrm{t}\mathrm{U}\mathrm{p}\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{d}={\mathrm{x}}_{\mathrm{c}\mathrm{u}\mathrm{t}}$); if too small, it moves to a larger position ( $\mathrm{S}\mathrm{e}\mathrm{t}Lowerbound=$ $x_{cut}$). When a position is found where the volume is close enough to the target, the position and point cloud are saved as $X_{answer}$ and $Part\mathrm{\_}X$, respectively. The process repeats for each vertical segment until all parts are defined, with the last segment covering everything up to $x_{max}$.

Horizontal Segmentation (Lines 36–64, Fig. 10): After dividing the object vertically, each vertical part is further divided horizontally along the y-axis. For each vertical part that is saved in $\mathrm{P}\mathrm{a}\mathrm{r}\mathrm{t}\mathrm{\_}\mathrm{X}$, the algorithm first identifies the object’s minimum and maximum y-coordinates ( ${\mathrm{y}}_{\mathrm{m}\mathrm{i}\mathrm{n}}$, ${\mathrm{y}}_{\mathrm{m}\mathrm{a}\mathrm{x}}$), indicated by the dashed lines in Fig. 10, left. It then searches for a cutting position ${\mathrm{y}}_{\mathrm{c}\mathrm{u}\mathrm{t}}$ (shown as the red line in Fig. 10, left) such that the volume between ${\mathrm{y}}_{\mathrm{m}\mathrm{a}\mathrm{x}}$ and ${\mathrm{y}}_{\mathrm{c}\mathrm{u}\mathrm{t}}$ closely matches the desired volume. If the volume is too large, the position is adjusted closer to a smaller position ( $\mathrm{S}\mathrm{e}\mathrm{t}\mathrm{L}\mathrm{o}\mathrm{w}\mathrm{e}\mathrm{r}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{d}={\mathrm{y}}_{\mathrm{c}\mathrm{u}\mathrm{t}}$); if too small, it moves to a larger position ( $\mathrm{S}\mathrm{e}\mathrm{t}\mathrm{U}\mathrm{p}\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{d}={\mathrm{y}}_{\mathrm{c}\mathrm{u}\mathrm{t}}$). When the volume is close enough to the target, the position and point cloud is saved as ${\mathrm{Y}}_{\mathrm{a}\mathrm{n}\mathrm{s}\mathrm{w}\mathrm{e}\mathrm{r}}$ and $\mathrm{P}\mathrm{a}\mathrm{r}\mathrm{t}\mathrm{\_}\mathrm{Y}$ respectively. This process is repeated until all horizontal lines for each vertical part are defined, with the last slice covering everything up to ${\mathrm{y}}_{\mathrm{m}\mathrm{i}\mathrm{n}}$.

Output (Lines 65–68): Finally, the algorithm outputs the positions of the cutting lines along the x-axis ( ${\mathrm{X}}_{\mathrm{a}\mathrm{n}\mathrm{s}\mathrm{w}\mathrm{e}\mathrm{r}}$) and y-axis ( ${\mathrm{Y}}_{\mathrm{a}\mathrm{n}\mathrm{s}\mathrm{w}\mathrm{e}\mathrm{r}}$), and the segmented parts ( $\mathrm{P}\mathrm{a}\mathrm{r}\mathrm{t}\mathrm{\_}\mathrm{X}$and $\mathrm{P}\mathrm{a}\mathrm{r}\mathrm{t}\mathrm{\_}\mathrm{Y}$) are stored for further use. For example, the cutting lines shown in Fig. 11 are desired, the values that must be set are ${\mathrm{n}}_{\mathrm{v}}=2$ and ${\mathrm{n}}_{\mathrm{h}}=\left[2,3\right]$.

After computing these cutting lines, the algorithm generates 3D intersection planes that represent the actual cutting surfaces. Each intersection is extracted and saved individually as a .PLY (Polygon File Format) file. In the evaluation phase described in ‘Cutting Algorithm Evaluation’, these PLY files are used to project cutting lines onto the sample surface via a laser engraving module.

Setup system for evaluation

To verify that the cutting lines generated by the proposed method could accurately divide the object into equal-weight portions, laser engraving was used to apply the cutting lines onto the object’s surface. This technique burned markings along the calculated lines, providing a clear visual representation for further analysis and evaluation. The setup consisted of three main components as shown in green area of Fig. 2: a cartesian robot, which enabled movement along the X, Y, and Z axes; a laser engraving module, mounted on the robot to mark the cutting lines on the meat surface; and a movable base plate (jig), which ensured accurate positioning between the machine vision system and the evaluation system.

Evaluation method

Many test samples were required to comprehensively evaluate the algorithm’s capability to divide objects into equal-weight portions. To minimize the loss of raw chicken meat and reduce experimental costs, modeling clay was selected as the primary test material during the initial evaluation stages. Clay samples allowed repeated trials under controlled conditions while maintaining consistent shape and weight. The evaluation process incorporated multiple statistical methods, including verification of the algorithm’s accuracy in object volume calculation, one-way ANOVA to assess the effect of different cutting patterns on segmentation accuracy, and Gage R&R analysis to quantify measurement system reliability across multiple segments. Following these preliminary tests, the algorithm was validated on actual chicken breast and fish fillets to confirm its practical applicability under realistic processing conditions.

Figure 12 illustrates the data collection process in preparation for the evaluation procedure, starting with scanning the object and determining cutting lines using the machine vision system. Once the computed cutting lines were identified, they were projected onto the object surface using the laser engraving module mounted on the cartesian robot. The system first received the cutting coordinates from the cutting algorithm and converted them into precise movement commands for the robot. The robot then positioned the laser over the designated cutting lines and initiated the engraving process. Figure 13A shows that the engraved cutting lines serve as guides for the subsequent cutting process, where operators used knives to follow the marked lines and divide the object accordingly. After the object was cut, each piece was weighed to collect data for further evaluation, with the sequence of each piece being recorded, as shown in Fig. 13B.

The experiments are as follows:

• (1)

  Segmentation errors and limitations in point cloud volume calculation

• This experiment was conducted to verify that the volume-segmentation algorithm partitions each object into n segments matching the target volume.

• (2)

  Testing different cutting patterns

• To evaluate how cutting configurations affect weight uniformity, we prepared modeling clay samples of 200 g and applied five cutting patterns. Each pattern was repeated five times (Total 25 samples). Cutting lines were generated by the algorithm and projected via laser for manual separation and weighing, following the data-collection workflow in Fig. 12. The primary outcomes were MAE and percentage MAE (%MAE), computed as in Eqs. (3), (4). Data screening included the Kolmogorov–Smirnov test for normality and Levene’s test for homogeneity. Group differences in MAE across cutting patterns were analyzed using one-way ANOVA. (3) $$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{\mathrm{n}}\sum _{\mathrm{i}=1}^{\mathrm{n}}\left|{\mathrm{y}}_{\mathrm{i}}-\mathrm{x}\right|,}$$ (4) $$\mathrm{\%}\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{\mathrm{M}\mathrm{A}\mathrm{E}}{\mathrm{x}}\times 100,}$$where:

  $\mathrm{M}\mathrm{A}\mathrm{E}$ = Mean Absolute Error

• ${\mathrm{y}}_{\mathrm{i}}$ = Measured weight of the object segment

• x = Expected weight per segment

• n = Number of segments the object is divided into

• A one-way ANOVA test was conducted to determine whether significant differences existed between the cutting patterns in terms of MAE. This analysis aimed to ascertain the extent to which different cutting configurations influenced segment weight consistency. Before conducting ANOVA, data normality was validated using the Kolmogorov-Smirnov test, while Levene’s test was applied to confirm the homogeneity of variance (Massey, 1951). Sample size (N) was determined with G*Power (Faul et al., 2007) for a one-way ANOVA. The study was structured with the following parameters: significance level (α) = 0.05, power = 0.8, and five experimental groups. The effect size (Cohen, 2013) used in this study (f = 0.852) was computed in G*Power using Eq. (5), based on group means obtained from a pilot study, leading to a minimum total sample size of 25 pieces, ensuring five repetitions per cutting pattern. (5) $$\mathrm{f}={\displaystyle \frac{\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{k}}{\mathrm{n}}_{\mathrm{i}}{\left(\overline{\mathrm{X}}-\overline{\mathrm{X}}\right)}^2/\left({\mathrm{n}}_1+{\mathrm{n}}_2+\dots +{\mathrm{n}}_{\mathrm{k}}\right)}}{\mathrm{\sigma }},}$$where:

  ${\mathrm{n}}_{\mathrm{i}}$ Sample size for each group

• $\overline{{\mathrm{X}}_{\mathrm{i}}}=$ Mean of each group

• $\overline{\mathrm{X}}=$ Grand mean

• $\mathrm{k}=$ Number of groups

• $\mathrm{\sigma }=$ Standard deviation within groups

• (3)

  Gage R&R (Gauge Repeatability and Reproducibility)

  Gage R&R is a statistical method used to evaluate the quality of a measurement system by assessing its repeatability (the consistency of measurements taken by the same instrument under the same conditions) and reproducibility (the consistency of measurements taken by different operators or across different groups). Typically, a Gage R&R study begins by selecting multiple test samples and multiple operators. Each operator then measures each sample multiple times to assess variability. In this thesis, Gage R&R was applied to analyze the weight of the first modeling clay piece after it had been cut.

• The data collection process involved multiple steps to ensure the accuracy and reliability of the weight measurements after the cutting process. First, 81 modeling clay samples of 100, 150, and 200 g were prepared and molded into uniform shapes to minimize external variations, as shown in Fig. 14. Each sample was then divided into 4, 5, or 6 pieces using algorithm-generated cutting lines. To simulate variability in operator placement commonly found in industrial environments, the samples were intentionally positioned in three different orientations before scanning. For example, the 100 g samples were rotated approximately 90° and 120° from their initial position, the 150 g samples were laterally shifted by ±20 mm, and the 200 g samples were rotated to 90° and 180°, with minor displacements up to ±5 mm and angular variations up to ±10°. Weight measurements were collected for the first three segments (Segment 1–3) of each sample to evaluate the consistency and reliability of the system across different portions of the object. Each segment was weighed using a precision scale after manual cutting along the engraved paths, and the data were recorded in a structured sequence.

• 4)

  Real meat portioning test

• To validate the performance of the developed cutting algorithm on actual food products, experiments were conducted using chicken breast and fish fillets as test materials. The aim was to evaluate the algorithm’s capability to portion real meat into equal-weight segments.

• For each meat type, 30 individual samples were used (total of 60 samples). The samples were randomly allocated into three cutting configurations:

  • •

    4 segments (chicken breast 10 samples, fish 10 samples)

  • •

    5 segments (chicken breast 10 samples, fish 10 samples)

  • •

    6 segments (chicken breast 10 samples, fish 10 samples)

• The initial weights of the meat pieces varied within each group to represent realistic processing conditions. The machine vision system scanned each sample to acquire point cloud data, and the cutting algorithm computed vertical and horizontal cutting lines to achieve equal-volume segmentation. The calculated cut positions were projected onto the meat using the laser marking system of the evaluation setup, after which manual cutting was performed along the marked lines. Following cutting, each segment was weighed. The primary performance metric was the percentage mean absolute error (%MAE) between the actual and target weights, calculated as Eqs 3 and 4. The standard deviation (SD) of segment weights and the maximum and minimum %MAE values within each group were also recorded.

One-way ANOVA for cutting pattern analysis

The analysis was based on the data presented in Table 2. Before conducting ANOVA, data normality was validated using the Kolmogorov-Smirnov test, while Levene’s test was applied to confirm the homogeneity of variance (Massey, 1951). Results indicated that the data set met the assumptions necessary for ANOVA.

The result of ANOVA test in Table 3 revealed statistically significant differences among at least one pair of cutting patterns, F(4, 20) = 6.029, p = 0.002.