Literature Reading 001

The title of the document is as follows:

[1] 2002 ICIP face photo recognition using sketch
[2] 2005 CVPR A nonlinear approach for face sketch synthesis and recognition
[3] Data-driven vs. Model-driven Fast face sketch Synthesis

Just went into the lab this week and read a cursory review of the three papers recommended by Mr. Zhang about the direction of the study.

1. General direction: Face Recognition (facial recognition)
2. Small direction: Based on the sketch image method (based on sketch)

background :
In the past, face recognition was based on the original image (based on photo), but when a suspect was to be identified and no photo was taken, a sketch image (sketch) was replaced by the eyewitness description. Therefore, you need to convert the original image taken to a sketch picture, and then compare it with the sketch gallery .

The main contents of the Thesis:

[1] 2002 ICIP face photo recognition using sketch

1. Traditional methods for extracting facial features (conventional eigenface method): KLT
2. Sketch-based approach (Eigensketch transformation)

[2] 2005 CVPR A nonlinear approach for face sketch synthesis and recognition

The paper is divided into two major tasks: transformation and recognition

1. Conversion: This paper is based on the LLE method, the model (P) trained with Photo-sketch pair training set is automatically generated Pseudo-sketch.
2. Recognition: Because the painter's drawings are inevitably biased, and pseudo-sketch will be blurred, so use the Knda method.

question : Lle,linear V.s Nonlinear and other knowledge points have not been understood.

[3] Data-driven vs. Model-driven Fast face sketch Synthesis

Data-driven:
Data-driven is trained at the levelof the module, including two models: Neighbor Selection model, Neighbor Fusion model (e.g.)
The authors cite a number of Data-driven methods, which illustrate their disadvantage: it takes too much time to traverse the entire training photo-sketch pairs when choosing neighbor Selection Model.

question : How is patch division done? --partition Mask (e.g.)

Model driven:
Training: Divide photo-sketch pairs into many patches and train many mapping functions.
Test: The photo is divided into many patches, using the trained mapping functions to generate the corresponding sketch.

Just a preliminary understanding of the direction, the specific principles, formulas, algorithms have not been thoroughly understood. Read in encountered some do not understand the point of knowledge directly skip, such as rigid regression, LLE, Knda and so on. Later, we will read and summarize each article further.

Literature Reading 001