Instant noodle anomaly detection based on few-shot learning and feature alignment

Abstract

ObjectiveTo achieve efficient detection of fine foreign objects in instant noodle, this paper proposes an instant noodle anomaly detection method based on few-shot learning and feature alignment.MethodsThe method uses a pre-trained residual network as a feature extraction network for efficiently extracting the features of the instant noodle image. The image is geometrically transformed by introducing a spatial transformation network for better alignment and extraction of features. Feature alignment is used to align the key features in the image so that the accuracy and generalization ability of the model can be ensured in detecting instant noodle anomalies.ResultsExperiments are conducted on a self-made instant noodle dataset, and the proposed method achieves areas under the curve (AUCs) of 86.2% and 93.3% at the 5-shot image level and pixel level, respectively, which outperforms other methods on the instant noodle dataset.ConclusionThe proposed method can effectively detect fine foreign objects in the instant noodle anomaly detection task.