Beef marbling assessment by structured-illumination reflectance imaging with deep learning

Beef marbling is primarily assessed in terms of the abundance and spatial distribution of visible fat flecks within Longissimus Dorsi (LD) muscle. Visual appraisal by trained professionals is the currently prevailing practice in beef marbling assessment, but it is subjective, labor-intensive, and time-consuming. Structured-Illumination Reflectance Imaging (SIRI) offers a potentially useful approach to meat quality assessment as opposed to conventional imaging techniques using uniform illumination. This study presents a new evaluation of a custom-assembled, broadband SIRI system, combined with Deep Learning (DL) approaches, for beef marbling assessment. Beef samples of varying marbling degrees were imaged under sinusoidal illumination at spatial frequencies of 0.05-0.40 cycles/mm. The acquired images were demodulated into Direct Component (DC) and Amplitude Component (AC) images at each spatial frequency. Three DL-based segmentation models, including Unet++, DeepLabv3+, and SegFormer, were built using DC and AC (0.05-0.40 cycles/mm) images for segmenting the LD muscle from the surrounding backfat. Two scenarios of image input, i.e., 1) original beef images without backfat removal and 2) segmented LD muscle images, were examined for feature extraction using a pre-trained ResNeXt-101 model, followed by discriminant modeling to classify beef samples of three marbling categories. SegFormer achieved better LD segmentation than Unet++ and DeepLabv3+ for both DC and AC images, with the best IoU (Intersection over Union) of 96.8% obtained at 0.10 cycles/mm. For marbling classification, the models based on LD segments (scenario 1) consistently produced higher accuracies than those using the original beef images (scenario 2). With the deep features extracted from LD segments, DC images yielded an overall classification accuracy of 83.19%, while the AC images resulted in better accuracies of 85.96%-88.72% (P<0.05), with the greatest improvement of 5.53% (P<0.01) at 0.40 cycles/mm. A further improved accuracy of 90.85% in the marbling classification was achieved by modeling a reduced set of relevant features identified based on MRMR (Minimum Redundancy Maximum Relevance). This study has shown the effectiveness of SIRI for beef marbling assessment.