Self-Checkout Product Class Verification using Center Loss approach

Abstract

The traditional image classifiers are not capable to verify if samples belong to specified classes due to several reasons: classifiers do not provide boundaries between in-class and out-of-class samples; although classifiers provide separation boundaries between known classes, classifiers' latent features tend to have high intra-class variance; classifiers often predict high probabilities for out-of-distribution samples; training classifiers on unbalanced data results in bias towards over-represented classes. The nature of the class verification problem requires a different loss function than the ubiquitous cross entropy loss in traditional classifiers: input to a class verification function includes a suggested class in addition to an image. As opposed to outlier detection, space is transformed to be not only separable, but discriminative between in-class and out-of-class inputs. In this paper, class verification based on a euclidean distance from the class centre is proposed and implemented. Class centres are learnt by training on a centre loss function. The method's effectiveness is shown on a self-checkout image dataset of 194 food retail products. The results show that a two-fold loss function is not only useful to verify class, but does not degrade classification performance - thus, the same neural network is usable both for classification and verification.