next up previous contents
Next: Conclusions Up: Testing Previous: Recognition Test

Sensitivity Analysis

Evidently, the recognition stage depends on the accuracy of the localization stage. We wish to observe the effect of localization errors on the recognition process to quantify the sensitivity of the recognition algorithm. Recall the normalization procedure described in Chapter 4. We shall represent this procedure as a function, N, which acts on an image, I, to produce a standard mug-shot (probe) which can be used to search for a match in the database. The parameters of the normalization are the (x,y) positions of the 4 feature points i0, i1, i2 and i3 (left eye, right eye, nose and mouth) and the original intensity image I. This computation is shown in Equation [*]:


 
probe = N(i0x,i0y,i1x,i1y,i2x,i2y,i3x,i3y,I) (5.1)

We then introduce the concept of recognition certainty. Certainty involves comparing the distance a face has from its correct match in the database to the distance it has from other members in the database. Evidently, our recognition output is more reliable (or has better certainty) if a face is much closer to its match and than it is to other members of the database. Consider face#0 from the database which is shown with its synthesized mug-shot image in Figure [*]. The eyes, nose and mouth have been localized accurately at positions i0', i1', i2' and i3'. These anchor points are used to generate a mug-shot version of face#0 using Equation [*].


  \begin{figure}% latex2html id marker 4428
\center
\begin{tabular}[b]{cc}
\eps...
... with anchor points.
(b) Corresponding synthesized mug-shot image.}\end{figure}

This face is a member of our database (face0) and so d(probe,face0)=d(face0,face0)=0. However, if we perturb the values i0', i1', i2' and i3' by a small amount, the resulting probe image using Equation [*] will be different and dmin will no longer be 0. The distance of the probe face (probe) to the correct match face#0 in the database is defined as d(probe,face0). The distance to the closest incorrect response is $d_{min \vert x\epsilon (0,P)}$, as defined in equation [*]:


 \begin{displaymath}d_{min \vert x\epsilon (0,P)}=min_{x=1}^{x<P} d(probe,face_x)
\end{displaymath} (5.2)

In Equation [*] we compute a margin of safety which is positive when the probe resembles its correct match face#0 and negative when the probe matches another element of the database, instead.


 \begin{displaymath}c=d_{min \vert x\epsilon (0,P)}-d(probe,face_0)
\end{displaymath} (5.3)

Since the probe is actually the result of the function N in Equation [*], it has i0, i1, i2 and i3 as its parameters as well. By the same token, the value of c in Equation [*] also has i0, i1, i2 and i3 as parameters. Thus, it is more appropriate to write c(i0x,i0y,i1x,i1y,i2x,i2y,i3x,i3y). However, for compactness, we shall only refer to the certainty as c.

We shall now compute the sensitivity of the certainty (c) to variations in the localization (i0, i1, i2 and i3) for the image Icorresponding to face#0. This is done by computing $c(i_{0_x}'+\Delta_{0_x},i_{0_y}'+\Delta_{0_y},i_{1_x}'+\Delta_{1_x},i_{1_y}'+\D...
...\Delta_{2_x},i_{2_y}'+\Delta_{2_y},i_{3_x}'+\Delta_{3_x},i_{3_y}'+\Delta_{3_y})$.

We vary the $\Delta$ values which cause the localization of a feature point to move around its original position. The anchor point's displacement caused by a particularly large $\Delta$ is depicted in Figure [*]. The dimensions of this image are $512\times342$ and the intra-ocular distance is approximately 60 pixels. In the experiments, the $\Delta$ values are varied over a range of [-15,15] pixels each. We then synthesize a new mug-shot image from the perturbed anchor points.


  \begin{figure}% latex2html id marker 4480
\center
\begin{tabular}[b]{cc}
\eps...
... with anchor points.
(b) Corresponding synthesized mug-shot image.}\end{figure}

Figure [*] shows several synthesized images after perturbing $\Delta_{0_x}$ and $\Delta_{0_y}$, with all other $\Delta$ values fixed at 0. Not surprisingly, the mug-shots that are synthesized appear slightly different depending on the position of i0 (the locus of the left eye). Similarly, Figure [*] shows the approximations after the KL encoding of the mug-shots in Figure [*]. The approximations, too, are affected, showing that the KL transformations is sensitive to errors in localization. Finally, in Figure [*] we show the value of c as we vary $\Delta_{0_x}$ and $\Delta_{0_y}$, with all other $\Delta=0$.


  
Figure 5.16: Synthesized mug-shots with left eye anchor point perturbed
\begin{figure}\center
\epsfig{file=implem/figs/mugs2.ps,height=14cm, angle=-90} \end{figure}


  
Figure 5.17: KL approximations to mug-shots with left eye anchor point perturbed
\begin{figure}\center
\epsfig{file=implem/figs/appshot2.ps,height=14cm, angle=-90} \end{figure}


  
Figure 5.18: Variation in recognition certainty under left eye anchor point perturbation
\begin{figure}\center
\epsfig{file=implem/figs/sdim.leye.ps,height=8cm} \end{figure}

Figure [*] shows the value of c for variations in the (x,y)position of the right eye anchor point. Similarly, Figure [*] and Figure [*] show the same analysis for the nose point under two different views. Finally, Figure [*] shows the effect of perturbing the mouth point. This surface is quite different from the ones in the previous experiments. In fact, the value of c stays constant and positive indicating that the changes in the mouth position have brought no change in the synthesized mug-shot and that the recognition performance is unaffected by mouth localization errors. This is due to the insensitivity the 3D normalization procedure (defined in Chapter 4) has to the locus of the mouth point.


  
Figure 5.19: Variation in recognition certainty under right eye anchor point perturbation
\begin{figure}\center
\epsfig{file=implem/figs/sdim.reye.ps,height=8cm} \end{figure}


  
Figure 5.20: Variation in recognition certainty under nose anchor point perturbation (View 1)
\begin{figure}\center
\epsfig{file=implem/figs/sdim.nose.ps,height=8cm} \end{figure}


  
Figure 5.21: Variation in recognition certainty under nose anchor point perturbation (View 2)
\begin{figure}\center
\epsfig{file=implem/figs/sdim.nose2.ps,height=8cm} \end{figure}


  
Figure 5.22: Variation in recognition certainty under mouth anchor point perturbation
\begin{figure}\center
\epsfig{file=implem/figs/sdim.mouth.ps,height=8cm} \end{figure}

The above plots show that the localization does not have to be perfect for recognition to remain successful. In these graphs, as long as the value of cis positive, then the subject is identifiable. Even though these sensitivity measurements were made by varying only two parameters, the 8 dimensional sensitivity surface can be approximated by a weighted sum of the 4 individual surfaces described above. Thus, an 8 dimensional sub-space exists which defines the range of the 8 $\Delta$ perturbations that will be tolerated before recognition errors occur. In short, the anchor-point localizations may be perturbed by several pixels before recognition degrades excessively.

From the above plots, it is evident that the $\Delta_{2_y}$ (the nose locus) is the most sensitive anchor point since it causes the most drastic change in c. Consequently, an effort should be made to change the normalization algorithm to reduce the sensitivity of the recognition to this locus. On the other hand, there is a large insensitivity to the location of the mouth. This is due to the limited effect the mouth has in the normalization procedure. In fact, the mouth only determines the vertical stretch or deformation that needs to be applied to the 3D model. Thus, an effort should be made to use the location of the mouth more actively in the normalization algorithm discussed in Chapter 4. Recall that an error Emouth was present in the normalization while the other 3 anchor points always aligned perfectly to the eyes and nose in the image (using the WP3P). Thus the 3D normalization has an alignment error concentrated upon the mouth-point which is the only point on the 3D model which does not always line-up with its destination on the image. If this error could be distributed equally among all four features points, each point will be slightly misaligned and the total misalignment error would be less. Consequently, the 3D model's alignment to the face in the image would be more accurate, overall. Thus, we would attempt to minimize Etotal as in Equation [*] instead of minimizing Emouthwith Eleft-eye=Eright-eye=Enose=0. The end result would be an overall greater insensitivity and recognition robustness for the 8 localization parameters combined.


 \begin{displaymath}E_{total}=\sqrt{E_{mouth}^2+E_{left-eye}^2+E_{right-eye}^2+E_{nose}^2}
\end{displaymath} (5.4)

We have thus presented the system's structure as a whole. The localization and recognition tests evaluate its performance. For one training image, the system is competitive when compared to contemporary face recognition algorithms such as the one proposed by Lawrence[23]. Other current algorithms include [32] and [29] which report recognition rates of 98% and 92% respectively. However, these algorithms were tested on mostly frontal images (not the Achermann database or similar database). Finally, a sensitivity analysis depicts the dependence of the recognition module on the localization output. We see that the localization does not have to be exact for recognition to be successful. However, the sensitivity plots do show that recognition is not equally sensitive to perturbations in the localization of different anchor point. Thus, the normalization process needs to be adjusted to compensate for this discrepancy.


next up previous contents
Next: Conclusions Up: Testing Previous: Recognition Test
Tony Jebara
2000-06-23