Deep learning-based acoustic sorting of shijimi clams: performance across recording conditions and false-negative risk
Year: 2026
Type: article
Abstract: False-negative errors are critical in acoustic food inspection because non-acceptable samples may be misclassified as acceptable. We evaluated a deep learning-based acoustic sorting workflow for shijimi clams ( Corbicula japonica ), emphasizing NG recall and false-negative rate (FNR) rather than overall accuracy. Here, “risk-aware” denotes an application-oriented evaluation workflow rather than a new learning algorithm. Impact sounds were converted into MFCC feature maps, and seven CNN architectures were compared using fixed ten-fold cross-validation. Several architectures performed comparably; EfficientNet-B0 was retained according to a predefined NG-prioritized ranking rule, without claiming statistical superiority. In a same-dataset benchmark with identical fold assignments, an RBF SVM performed similarly (NG recall, 0.987 ± 0.011; FNR, 0.013 ± 0.011). In a temporally distinct clam batch collected in June 2026, the model achieved correct classification rates of 98.3% for NG and 99.4% for OK samples under the reference condition without retraining. Performance remained high across clam-size classes and between two microphone units, but increasing recording volume or reducing release-to-impact distance substantially degraded classification. Because hyperparameter screening and cross-validation used the same primary dataset, performance estimates may be mildly optimistic. This controlled proof-of-concept requires external validation across independent batches, seasons, devices, clam sources, and production-line environments.
Source: Applied Food Research
Institution: Shimane University
Language: English
Subfield: Analytical Chemistry
Field: Chemistry
Domain: Physical Sciences
SDG: Zero hunger
Open Access status: gold
Estimated APC paid: $2,050
Funder: Shimane University
OpenAlex ID: W7204209670