Method 1: directly find the point that is smaller than its neighbours

Cons: seems to be hard cut-off and not very flexible

1. use matlab function  BW = imregionalmax(I,conn), where conn defines the connected neighbours around a local minimum point, conn can be 4 or 8 in 2D situation  http://au.mathworks.com/help/images/ref/imregionalmax.html#inputarg_conn  

Comment: I think it is a little bit lack of flexibility, for example, what if the size of local minimum changes in a wide range and thus even 8 connected neighbours are not enough. Or what if the noisy pixel is in the neighbours that the center point is bigger than the nosie?

2 Raphael's method: based on a paper, energy function optimization method

2. Raphael's idea: first find the local minimum point (similar to using imregionalmax), then delet the noisy points

Method 2  Boundary-based method/ image segmentation method

Idea: find the boundary of a local cluster, which contains a local minimum.

1. bw = activecontour(A,mask,method)

Comment: snake contour and active contour are similar

                 Can we try using watershed algorithm, which is also a dilation algorith,?

2. errosion and expasion

Method 3 signal processing methods: for example, detect valleys in 1-D/2D signals

Idea: The common way is to first smooth the whole picture, so as to get rid of noise, and then find local minimum.

Cons: But after smoothing, the location of local minimum in the smoothed map will probably be differnt from the locations in the original map

1. http://www.mathworks.com/matlabcentral/fileexchange/37388-fast-2d-peak-finder 

Cons: This algorithm only consider 1-pixel noise, which is not true

2. http://blogs.mathworks.com/pick/2008/05/09/finding-local-extrema/  find valleys in a 1D signal

P.S. refer to http://au.mathworks.com/matlabcentral/newsreader/view_thread/102944 for more answers