| Title |
Parameter-efficient MobileNetV2 subnetwork selection for facial expression recognition in low data regimes |
| Translation of Title |
Parametrams taupi MobileNetV2 poaibių parinktis veido išraiškų atpažinimui, esant mažiems duomenų kiekiams. |
| Authors |
Thuruthel Murali, Ananthakrishnan |
| Full Text |
|
| Pages |
70 |
| Keywords [eng] |
facial expression recognition ; transfer learning ; mobilenetv2 ; subnetwork selection ; small dataset |
| Abstract [eng] |
Facial expression recognition is becoming an important topic with the introduction of self-driving cars, or whether it be to find the mental state of a patient in a hospital or to find out if a person is lying. A key challenge in this field is achieving reliable performance under small dataset conditions, where standard deep learning methods are prone to overfitting. This study investigated the use of MobileNetV2 based subnetwork selection as a parameter efficient approach to FER under limited data conditions. A global search strategy was developed to identify the optimal subnetwork, defined as blocks 1 through k, where k ∈{3…..17}, by evaluating the candidate paths using a lightweight logistic regression probe on a subject independent subset of the data. The selected subnetwork was then finetuned and evaluated against a full MobileNetV2 baseline across FER2013 and CK+ datasets. The exhaustive search was able to identify blocks 1-13 as the optimal subnetwork, achieving a 55.8% parameter reduction (2.23M to 0.99M). On FER2013, the subnetwork produced no statically significant accuracy at 10-20% data fractions while reducing the train-validation accuracy gap by 35-48% across all evaluated data fractions across all evaluated data fractions, which confirms a significant lower overfitting than the full model. On the CK+ dataset under subject independent five-fold cross validation, the subnetwork achieved a 90.47% accuracy compared to 88.58% for the baseline, with macro F1 improving from 0.834 to 0.862. These results indicate that a carefully selected compact subnetwork can produce competitive accuracy in a moderate low data setting while giving significant advantage in parameter efficiency and overfitting resistance. |
| Dissertation Institution |
Kauno technologijos universitetas. |
| Type |
Master thesis |
| Language |
Lithuanian |
| Publication date |
2026 |
