A New Laser Profiling System for Online, Real-Time Detection of Broiler Breast Fillets with Woody Breast

Jiaming Zhang, Yuzhen Lu*
Department of Biosystems and Agricultural Engineering, Michigan State University
📄 Paper 💻 Code 📎 Supplementary

Abstract

Woody breast (WB) myopathy is a muscle quality defect of poultry breast meat that causes product downgrading or rejection, and significant economic losses for poultry industries worldwide. Current detection of WB in poultry processing plants relies on manual palpation and visual inspection, which is labour-intensive and subjective. Surface profilometry or three-dimensional (3D) vision techniques that measure surface topography of objects offer a potentially useful method for WB assessment and grading, since WB alters the shape of chicken breasts. This study presents an innovative, custom-built 3D laser profiling system for online, real-time detection of broiler breast fillets with WB through 3D reconstruction and machine learning. The system employed a line laser to scan samples at a rate of 120 frames per second (fps), and with a dedicated calibrated algorithm pipeline, was capable of reconstructing the shape of samples at a rate of about 107 fps. Compared to a red line laser (λ=660 nm), a blue line laser (λ=450 nm) yielded better 3D reconstruction, with the z-axis (depth/height) reconstruction error of 0.29, 0.73, and 2.56 mm at the conveyor speed of 5, 10, and 15 cm/s, respectively; higher conveyor speeds resulted in reduced point cloud density and elevated image noise. A set of 310 chicken breast fillets, manually graded by trained personnel for WB conditions, was scanned under the illumination of a blue line laser at the three conveyor speeds for WB assessment. Classification models were built using two approaches, i.e., support vector machine trained with the hand-crafted features from the 2D projection of reconstructed shape, and deep learning through an end-to-end PointNet++ trained with the 3D points. At the conveyor speed of 5 cm/s, the PointNet++ model attained a better overall accuracy of 88.9%; the higher speed of 10-15 cm/s resulted in slightly reduced accuracy for both models. This study has demonstrated the promise of the proposed 3D laser profiling system for online, high-speed WB inspection of poultry meat, which has potential for practical application. The software programs of this study have been made publicly available.

Research Background

Woody Breast (WB) myopathy has emerged as one of the most costly quality defects in modern broiler production, with flock-level surveys showing overall incidences > 70% and ~ 12% of carcasses graded as "severe" [Petracci et al., 2015; Bordini et al., 2023; Che et al., 2022]. The resulting abnormal hardening of the Pectoralis major markedly lowers consumer acceptance and is estimated to cost the North-American poultry sector hundreds of millions of dollars annually (Kuttappan et al., 2016).

On today's processing lines, WB assessment still depends on manual palpation and visual inspection, which is labor-intensive, subjective, and exposes product surfaces to cross-contamination (Livingston et al., 2019; Pang et al., 2020; Ross et al., 2020). Reported mis-classification rates can reach 25%[1], and operator accuracy may fall to ≈ 60%[2] after prolonged shifts because of fatigue. These shortcomings create an urgent demand for an objective, rapid, and non-contact on-line solution.

Samples

Figure 1: (Left) Sampling and Woody Breast (WB) Assessment at a Poultry Processing Facility, and (Right) top view of normal and WB-affected samples.

In our earlier work we first proposed broadband light-scattering imaging (B-LSI), in which a collimated broadband beam illuminates the fillet and a monochrome camera records saturation/attenuation profiles; shape descriptors from these profiles achieved 92.3% accuracy in differentiating normal and WB fillets (Cai & Lu 2025a). We then built a hyperspectral LSI (H-LSI) prototype covering 400–1000 nm (Mu & Lu 2025). Both LSI variants rely on push-broom line scanning, requiring ~ 5 s (B-LSI) and ~ 12 s (H-LSI) per fillet—far slower than the sub-second cycle times demanded by commercial conveyors. To capture the ridge-like bulging characteristic of WB, we next developed structured illumination reflectance imaging (SIRI) with phase-measuring profilometry; dense surface geometry alone separated WB from normal fillets with ≈ 93% accuracy, an 8–11% gain over 2-D intensity features (Cai & Lu 2025b). SIRI, however, requires three phase-shift exposures per view, limiting belt speed and motivating our pursuit of high-throughput laser triangulation.

Method Overview

Active laser-triangulation ranging delivers sub-millimetre depth at video rates and remains robust on glossy, texture-less chicken surfaces. This project will therefore develop a line-laser triangulation 3-D vision system, integrating support-vector machines and a PointNet++ network, to enable real-time WB detection and grading on moving conveyor belts—bridging the gap between laboratory prototypes and the industrial need for high-throughput, reliable inspection.

Pipeline

Figure 2: The schematic of a 3D laser profiling system designed for online woody breast inspection of chicken breast fillets.

Figures 2 and 3 depict the overall workflow of system calibration and 3D reconstruction. Briefly, the calibration phase established the system's geometric reference frame, performed camera calibration, estimated the pose of each calibration board, and determined the laser plane. Meanwhile, the displacement of the calibration board was tracked, and the exact amount of movement of the conveyor belt between adjacent frames was recorded. With the knowledge of the laser plane, the detected pixels were to be converted instantly and rigidly into a dense, single-frame point cloud and merged with previous clouds to form a global model in the 3D reconstruction phase. After further point cloud processing (e.g., denoising and fillet segmentation), the consolidated point cloud would be utilized for downstream modeling and sample classification.

Workflow

Figure 3: System calibration and online 3D reconstruction workflow of the laser profiling system.

We employ two complementary pipelines for classifying WB point clouds:

1. Manual features + SVM — First, the 3D point cloud is projected into a 192×192 height map, then texture features (LBP, BSIF, and HOG) are extracted, standardized, and input into RBF-SVM, with parameters optimized through cross-validation.

2. Enhanced PointNet++ — Center-normalize the original point cloud, add normal vectors and curvature, and use four-level MSG set-based sampling to learn local-global geometric features, then directly output the binary classification probability.

Both models were tested on three speed conditions (5, 10, and 15 cm/s) with training, validation, and test sets randomly split into 70/15/15% and repeated 50 times.

Experimental Results

× Detail view
× Detail view
Table 1 · Reconstruction Performance for a Trapezoidal Calibration Block (TCB)
Conveyor speed Mean Absolute Error (MAE, mm) Mean Absolute Percentage Error (MAPE, %) Density (PTS/cm2)
x-axis y-axis z-axis x-axis y-axis z-axis
5 cm/s 1.58 0.52 0.29 1.50 0.68 0.87 82.7
10 cm/s 1.75 3.44 0.73 1.66 4.52 2.21 46.6
15 cm/s 2.13 5.30 2.56 2.03 6.96 7.70 31.0
Average 1.82 3.09 1.19 1.73 4.05 3.59 53.4
Side-view visualization

Figure 4: Side-view (Y-Z) of 3D reconstruction of a trapezoidal calibration block (TCB) (top), normal fillet (middle), and WB-affected fillet (bottom) at three conveyor speeds, where the color indicates height on the z-axis.

Top-view visualization

Figure 5: Top-view (X-Y) of 3D reconstruction of a trapezoidal calibration block (TCB) (top), normal fillet (middle), and WB-affected fillet (bottom) at three conveyor speeds, where the color indicates height on the z-axis.

Table 2 · WB Classification Performance Metrics (50× Repeats)
Conveyor speed Feature-based SVM PointNet++
ACC (%) Recall (%) ACC (%) Recall (%)
5 cm/s 88.42 ±2.03 87.66 ±2.73 88.91 ±1.83 89.5 ±2.11
10 cm/s 86.19 ±4.12 85.12 ±3.58 84.28 ±3.87 83.0 ±3.74
15 cm/s 82.53 ±4.68 81.01 ±4.89 80.91 ±4.13 79.0 ±4.40

Citation

@unpublished{zhang2025laser,
    title   = {A New Laser Profiling System for Online, Real-Time Detection of
              Broiler Breast Fillets with Woody Breast},
    author  = {Zhang, Jiaming and Lu, Yuzhen},
    journal = {Journal of Food Engineering},
    year    = {2025},
    note    = {Under review}
}

Acknowledgements

This work is supported by the Michigan Alliance for Animal Agriculture (M-AAA), project ID number AA-23-0029, from AgBioResearch and MSU Extension at Michigan State University, in partnership with the Michigan Department of Agriculture and Rural Development. The authors thank Miller Poultry (Orland, IN) for providing chicken fillets and assisting in manual assessment of WB conditions in this study.