Progressive Kernel Extreme Learning Machine for Food Image Analysis via Optimal Features from Quality Resilient CNN

Recently, food recognition has received more research attention for mHealth applications that use automated visual-based methods to assess dietary intake. The goal is to improve the food diaries by addressing the challenges faced by existing methodologies. In addition to the classical challenge of the absence of rigid food structure and intra-class variations, food diaries employing deep networks trained with pristine images are susceptible to quality variations in real-world conditions of image acquisition and transmission. Similarly, existing progressive classifiers that use visual features via a convolutional neural network (CNN) classify food categories and cannot detect food ingredients. We aim to provide a system that selects the optimal subset of features from quality resilient CNNs and subsequently incorporates the parallel type of classification to tackle such challenges. The first progressive classifier recognizes food categories, and its multilabel extension detects food ingredients. Following this idea, after extracting features from the quality resilient category and ingredient CNN models by fine-tuning it on synthetic images generated using the novel online data augmentation method random iterative mixup. Our feature selection strategy uses the Shapley additive explanation (SHAP) values from the gradient explainer to select the best features. Then, novel progressive kernel extreme learning machine (PKELM) is exploited to cater to domain variations due to quality distortions, intra-class variations, and so forth, by remodeling the network structure based on activity value with the nodes. PKELM extension for multilabel classification detects ingredients by employing a bipolar step function to process test output and then selecting the column labels of the resulting matrix with a value of one. Moreover, during online learning, the PKELM novelty detection mechanism can label unlabeled instances and detect noisy samples. Experimental results showed superior performance on an integrated set of measures for seven publicly available food datasets.