Like other lossless codecs for continuous-tone images, FELICS operates by
decorrelating the image and encoding it with an
entropy coder. The decorrelation is the context \Delta = H - L where H=max(P1,P2) and L=min(P1,P2) where P1,P2 are the pixel's two nearest neighbors (
causal, already coded and known at the decoder) used for providing the context to code the present pixel P. Except at the top and left edges, these are the pixel above and the pixel to the left. For example, the neighbors of pixel X in the diagram are A and B, but if X were at the left side, its neighbors would be B and D. P lies within the closed interval [L, H] roughly half the time. Otherwise, it is above H or below L. These can be encoded as 1, 01, and 00 respectively (p. 4). The following figure shows the (idealized) histogram of the pixels and their intensity values along the x-axis, and frequency of occurrence along the y-axis. The distribution of P within the range [L, H] is nearly uniform with a minor peak near the center (L+H)/2 of this range. When P falls in the range [L, H], P − L is encoded using an adjusted
binary code such that values in the center of the range use floor(log2(Δ + 1)) bits and values at the ends use ceil (log2(Δ + 1)) bits (p. 2). For example, when Δ = 11, the codes for P − L in 0 to 11 may be 0000, 0001, 0010, 0011, 010, 011, 100, 101, 1100, 1101, 1110, 1111. Outside the range, P tends to follow a
geometric distribution on each side (p. 3). It is encoded using a
Rice code with parameters chosen based on previous choices. For each Δ and each possible Rice code parameter
k, the algorithm keeps track of the total number of bits that would have been used to encode pixels outside the range. Then for each pixel, it chooses the Rice code with the based on Δ at the pixel. ==Improvements==