Abstract: A system and method facilitating document image compression utilizing a mask separating a foreground of a document image from a background is provided. The invention includes a pixel energy analyzer adapted to partition regions into a foreground and background. The invention further provides for a merge region component adapted to attempt to merge regions if the merged region would not exceed a threshold energy. Merged regions are partitioned into a new foreground and new background. Thereafter, a mask storage component stores the partitioning information in a binary mask.

Abstract: Systems and methods for encoding and decoding document images are disclosed. Document images are segmented into multiple layers according to a mask. The multiple layers are non-binary. The respective layers can then be processed and compressed separately in order to achieve better compression of the document image overall. A mask is generated from a document image. The mask is generated so as to reduce an estimate of compression for the combined size of the mask and multiple layers of the document image. The mask is then employed to segment the document image into the multiple layers. The mask determines or allocates pixels of the document image into respective layers. The mask and the multiple layers are processed and encoded separately so as to improve compression of the document image overall and to improve the speed of so doing. The multiple layers are non-binary images and can, for example, comprise a foreground image and a background image.

Abstract: A system and method facilitating image retouching is provided. The invention includes an image retoucher having a boundary detector and an image extender. The invention provides for the image retoucher to extend care pixels of at least one of a foreground and a background near a detected spurious boundary by altering the binary mask used for compression of the foreground and/or the background.

Abstract: A system and method facilitating activity (e.g., dithering/half toning and/or noise) detection is provided. The invention includes an activity detection system having a connected component analyzer and an activity detector. The invention provides for the quantity of connected component(s) in and/or intersecting a region surrounding a pixel to be determined. The activity detector provides an activity map output based, at least in part, upon the quantity of connected component(s) in and/or intersecting the region. The invention further provides for an optional image processor. In one example, if the quantity exceeds a first threshold, dithering/half toning is detected and appropriate action can be taken. Additionally, if the quantity is less than a second threshold, noise is detected and appropriate action can be taken.

Abstract: A system and method facilitating compression of bi-level images with explicit representation of ink clusters is provided. The present invention includes a cluster shape estimator that analyzes connected component information, extracts clusters and stores the cluster in a global dictionary, a page dictionary or a store of unclustered shapes. A bitmap estimation from clusters component determines dictionary positions for clusters stored in the global dictionary which are then encoded. A cluster position estimator determines page positions of clusters of the global dictionary and/or the page dictionary that are then encoded. Further, the global dictionary, the page dictionary and the store of unclustered shapes are also encoded.

Abstract: Systems and methods for performing adaptive filtering are disclosed. The present invention generates probabilities that can be used in an encoder, such as an arithmetic encoder and generates those probabilities in a computationally efficient manner. Probabilities of previously encoded coefficients are employed, effectively, in generating probabilities of the coefficients without regard to directional information. Thus, a large amount of information is adaptively and efficiently used in generating the probabilities. For the coefficients, the probability is computed based at least partly on at least one probability of a previously computed probability of a neighboring coefficient. Then, the coefficients are encoded using those computed probabilities.

Abstract: Compression of images that have masked or “don't care” regions which are delineated by a binary image mask is achieved using “masked wavelet transforms.” A unique mask-dependent lifting scheme is used to compute invertible wavelet transforms of the input image for use in encoding and decoding the input image. These mask-dependent wavelet transforms are derived from the input image based on the masked regions within the image. Masked wavelet coding automatically generates an appropriate linear combination of available, unmasked, neighboring pixels, for both the prediction and the update steps of “lifting” for each pixel. This pixel availability is then used to change the wavelet function on a case-by-case basis as a function of the mask by using a polynomial of degree k−1 for interpolation in both the predict and update steps of lifting where at least k unmasked neighboring pixel values are available.

Abstract: A system and method facilitating pattern recognition is provided. The invention includes a pattern recognition system having a convolutional neural network employing feature extraction layer(s) and classifier layer(s). The feature extraction layer(s) comprises convolutional layers and the classifier layer(s) comprises fully connected layers. The pattern recognition system can be trained utilizing a calculated cross entropy error. The calculated cross entropy error is utilized to update trainable parameters of the pattern recognition system.

Abstract: A system and method for performing trainable nonlinear prediction of transform coefficients in data compression such that the number of bits required to represent the data is reduced. The nonlinear prediction data compression system includes a nonlinear predictor for generating predicted transform coefficients, a nonlinear prediction encoder that uses the predicted transform coefficients to encode original data, and a nonlinear prediction decoder that uses the predicted transform coefficients to decode the encoded bitstream and reconstruct the original data. The nonlinear predictor may be trained using training techniques, including a novel in-loop training technique of the present invention. The present invention also includes a method for using a nonlinear predictor to encode and decode data. The method also includes improving the performance of the nonlinear prediction data compression and decompression using several novel speedup techniques.

Abstract: In order for neural network technology to make useful determinations of the identity of letters and numbers that are processed in real time at a postal service sorting center, it is necessary for the neural network to "learn" to recognize accurately the many shapes and sizes in which each letter or number are formed on the address surface of the envelope by postal service users. It has been realized that accuracy in the recognition of many letters and numbers is not appreciably sacrificed if the neural network is instructed to identify those characteristics of each letter or number which are in the category "invariant." Then, rather than requiring the neural network to recognize all gradations of shape, location, size, etc. of the identified invariant characteristic, a generalized and bounded description of the invariant segments is used which requires far less inputting of sample data and less processing of information relating to an unknown letter or number.

Abstract: Articles being graphical input patterns are sorted by a pattern-recognition machine that includes a data base of prototype patterns, each labeled with its respective class. The sorting method includes stops of storing input patterns in the pattern-recognition machine and classifying the stored input patterns. The classification is performed by calculating at least two distance functions between input patterns and prototype patterns. The distance functions belong to a hierarchy of distance functions that vary in the degree to which they are computationally intensive, and concomitantly, in their accuracy. Overall computational requirements are reduced by using less computationally intensive distance functions to preliminary filter out those prototype patterns that are farthest from the input pattern.

Abstract: Nodal outputs are discretized to values of S2.sup.n where n is an integer and S is equal to +1 or -1. During forward propagation, this offers the advantage of forming a product of a nodal output and a weight using a simple shift operation rather than a multiply operation. Replacing multiply operations with shift operations through out a neural network improves response times and permits building larger networks that have broader applicability. Training is also improved by increasing the efficiency of backward propagation. The multiplications involved in backward propagation are reduced To shift operations by discretizing the errors associated with each node so that they are represented as 2.sup.n where n is an integer and S is equal to +1 or -1.

Abstract: Speed and accuracy of the recognition process for an alphanumeric automated recognizer is enhanced by a novel scheme for comparing known prototype symbols to examples of unknown symbol. The scheme is invariant with respect to a selected set of small transformations of either the prototypes or the examples, requires only a few operators, and generates accuracy rates comparable to a human operator. The small transformations of interest are expressed locally by a process in which, during the data processing, the derivative of the transformed image with respect to the parameter that controls the transformation is calculated. This directional derivative then is used as the generator of interest in the recognition machine and is incorporated into the comparison process. The derivatives are generated by using tangent plane concepts, which yield a useful approximation of an otherwise not quantifiable complex surface in the neighborhood occupied by the unknown example.

Abstract: Trajectories are generated in response to an input label by using a reverse time delay neural network. The reverse time delay neural network comprises an input layer, a plurality of hidden layers, and an output layer, all arranged in succession so that the number of frames per layer increases as the network is traversed from the input layer to the output layer. Additionally, the number of features decreases as the network is traversed from the input layer to the output layer. Features of the trajectory are created from the input label so that a time series of frames can be output by the network. Frames generally relate to particular epochs of time or time units and a frame includes a plurality of features.Interconnection between layers is accomplished using differential neuron units. In the differential neuron unit, standard neuron weighting, summing, and nonlinear squashing functions are performed on the inputs thereto.