TABLE 4.
List of image-based food-recognition systems for dietary assessment based on hand-crafted features and “shallow” classifiers on publicly available food datasets1
| Reference | User input-preprocessing | Segmentation | Feature extraction | Dimensionality reduction | Classification | Volume estimation | Datasets and performance |
|---|---|---|---|---|---|---|---|
| (28) | Depth camera | SIFT, LBP, Gabor, and color features | SVM | — | NTU-FOOD ACC = 62.7% | ||
| (26) | Bag of Textons, PRICoLBP, and SIFT | Near duplicate image retrieval | — | UNICT-FD889 MAP = 67.5% | |||
| (30) | k-Means clustering | SURF, shape, and color features | Borda count scheme | — | Ambient Kitchen Precision = 86.29% Recall = 83.61% | ||
| (118) | Superpixel Linear Distance Coding and Locality-constrained Linear Coding, mid-level features | SVM | — | PFID ACC = 50.45%UEC-Food 100 ACC = 60.50% | |||
| (66) | Superpixels segmentation | Mid-level food parts approach | SVM | — | UEC-Food 100 ACC = 60.50% | ||
| (102) | SIFT, SURF | BOF | SVM | — | UEC-Food 100 ACC = 82.38% | ||
| (119) | Metric forests | — | Food-101 ACC = 68.29% | ||||
| (27) | SIFT, PRICoLBP, Textons, anti-Textons | BOF | ANN with χ2distance | — | UNICT-FD1200 ACC = 93.04% | ||
| (96) | Wavelet kernel-based Wu-and-Li Index Fuzzy clustering | Whale Levenberg Marquardt ANN | — | UNIMIB 2016 ACC = 96.27% | |||
| (99) | Multiple hypothesis segmentation: salient region detection, multi-scale segmentation and fast rejection | Color, texture and local neighborhood pixel features | ANN | — | UNIMIB 2016 ACC = 95.9% | ||
| (100) | Canny edge detection, multi-scale segmentation, fast rejection of background pixels | Color, texture, SIFT, and SURF features | 3-Layer ANN | — | UNIMIB 2016 ACC = 94.5% | ||
| (98) | Local variation segmentation | Color, texture, local descriptors: SIFT and Multi-scale Dense SIFT (MDSIFT) | Multi-kernel SVM | — | UNIMIB 2016 ACC = 93.9% | ||
| (33) | Color features, HOG, SIFT, LBP, Locality-constrained Linear Coding | Bag-of-Words | SVM | — | Menu-Match Recall = 83%Calorie estimation Absolute error = 232 ± 7.2 | ||
| (35) | Plate detection with Hough transform | Color and Edge Directivity Descriptor (CEDD), Gabor features, LBP, Local Color Contrast Chromaticity moments, Complex Wavelet features | KNN | — | UNIMIB 2015 (CEDD) ACC = 99.05% |
1
ACC, Accuracy; ANN, artificial neural network; BOF, bag-of-features; HOG, histogram of oriented gradients; KNN, k-nearest neighbors; LBP, local binary patterns; MAP, mean average precision; PFID: Pittsburgh fast-food image dataset; PRICoLBP, pairwise rotation invariant co-occurrence local binary patterns; SIFT, scale-invariant feature transform; SURF, speeded up robust features; SVM, support vector machine.