Added apply_random_color_offset()

dlib/image_transforms/random_color_transform.h

@@ -92,6 +92,63 @@ namespace dlib
         }
     }
 
+// ----------------------------------------------------------------------------------------
+
+    template <typename image_type>
+    void apply_random_color_offset (
+        image_type& img_,
+        dlib::rand& rnd
+    )
+    {
+        // Make a random color offset.  This tform matrix came from looking at the
+        // covariance matrix of RGB values in a bunch of images.  In particular, if you
+        // multiply Gaussian random vectors by tform it will result in vectors with the
+        // same covariance matrix as the original RGB data.  Also, this color transform is
+        // what is suggested by the paper:
+        //  Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet
+        //  classification with deep convolutional neural networks." Advances in neural
+        //  information processing systems. 2012.
+        // Except that we used the square root of the eigenvalues (which I'm pretty sure is
+        // what the authors intended).
+        matrix<double,3,3> tform;
+        tform = -66.379,    25.094,   6.79698, 
+                -68.0492, -0.302309,  -13.9539,
+                -68.4907,  -24.0199,   7.27653; 
+        matrix<double,3,1> v;
+        v = rnd.get_random_gaussian(),rnd.get_random_gaussian(),rnd.get_random_gaussian();
+        v = round(tform*0.1*v);
+        const int roffset = v(0);
+        const int goffset = v(1);
+        const int boffset = v(2);
+
+        // Make up lookup tables that apply the color mapping so we don't have to put a
+        // bunch of complicated conditional branches in the loop below.
+        unsigned char rtable[256];
+        unsigned char gtable[256];
+        unsigned char btable[256];
+        for (int i = 0; i < 256; ++i)
+        {
+            rtable[i] = put_in_range(0, 255, i+roffset);
+            gtable[i] = put_in_range(0, 255, i+goffset);
+            btable[i] = put_in_range(0, 255, i+boffset);
+        }
+        
+        // now transform the image.
+        image_view<image_type> img(img_);
+        for (long r = 0; r < img.nr(); ++r)
+        {
+            for (long c = 0; c < img.nc(); ++c)
+            {
+                rgb_pixel temp;
+                assign_pixel(temp, img[r][c]);
+                temp.red   = rtable[temp.red];
+                temp.green = rtable[temp.green];
+                temp.blue  = rtable[temp.blue];
+                assign_pixel(img[r][c], temp);
+            }
+        }
+    }
+
 // ----------------------------------------------------------------------------------------
 
 }

dlib/image_transforms/random_color_transform_abstract.h

@@ -71,7 +71,24 @@ namespace dlib
 
 // ----------------------------------------------------------------------------------------
 
-}
+    template <typename image_type>
+    void apply_random_color_offset (
+        image_type& img,
+        dlib::rand& rnd
+    );
+    /*!
+        ensures
+            - Picks a random color offset vector and adds it to the given image.  The offset
+              vector is selected using the method described in the paper:
+                Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet
+                classification with deep convolutional neural networks." Advances in neural
+                information processing systems. 2012.
+              In particular, we sample an RGB value from the typical distribution of RGB
+              values, assuming it has a Gaussian distribution, and then divide it by 10.
+              This sampled RGB vector is added to each pixel of img.
+    !*/
+
+// ----------------------------------------------------------------------------------------
 
 #endif // DLIB_RANDOM_cOLOR_TRANSFORM_ABSTRACT_Hh_