Holistic Siamese Model Optimized for Aged Face-Sketch Similarity Detection

The proposed work presents an Holistic approach for Face-sketch matching using a Siamese scenario. Proposed scheme is simple yet effective in overcoming the computational complexity of the feature based scheme and also provides a F1 score of 0.73 and accuracy of nearly 74%. The results are in demanding circumstances of using ageing effects in sketches. Optimum model from Siamese linkages is acquired with augmented sketches on CUHK dataset. The sketches are trained under augmented versions of zooming, distortion and cropped versions of them. The validation curve with merely L1 regularization peaks at 73% while it finds one at 77% with a combined L1 and L2 regularization for an average batch size of eight. The efficacy of the regularization in conjunction with augmentation prevents overfitting and provides effective results in ageing sketches.