SYSTEM AND METHOD FOR CONVERTING A COLOR FORMULA USING AN ARTIFICIAL INTELLIGENCE BASED CONVERSION MODEL BACKGROUND OF THE INVENTION 1. Field of the Invention [0001] The present invention relates generally to converting a color formula in one paint or color system to a formula in another paint or color system, and more particularly, to a method and system for converting a color formula using artificial intelligence.
2. Description of the Related Art [0002] Products today are offered to consumers in a wide variety of colors.
Consumer products may be colored by means of colorants, dye or paint. Color matching is required in a variety of areas, including textiles, plastics, various synthetic materials, prosthetics, dental applications, and paint applications, due to the many variations in color, due to the wide variations in shades and hues of any given color and color variations in an article. The actual color produced in a given article may vary due to a number of factors. For example, textile colors vary according to fiber composition. Colorants for plastic vary according to the plastic composition. Painted articles vary in color depending on any number of factors, such as paint composition, variations in the paint application process, including application method, film thickness, drying technique and number of layers. An important application for color matching is in the area of automotive color matching. Frequent uses for color matching in automotive paint occur in matching the same color from different batches or
matching similar colors from different manufacturers. Additionally, there is a requirement for color matching refinish paint to an OEM (original equipment manufacture) color when a vehicle body panels are damaged and require repainting.
[0003] A paint manufacturer supplies one or more paint formulations for the original paint color to refinish paint shops. By supplying a plurality of formulations or variants for a particular color, the paint manufacturer accounts for those factors which affect the actual color. Matching of dyes or colorants for other applications is also done through formulations for a particular color.
Typically, the formulations for a particular color are distributed on paper, microfiche, and/or compact disks (CD). A color tool, composed of swatches of the variants for each color may also be produced and delivered to each customer. The customer must select a formulation most closely matching the existing color of the article. This is typically done visually, i. e., by comparing swatches of paint or color to the part or in the case of paint, spraying a test piece with each formulation.
[0004] Different formulations are derived from actual data gathered by inspectors at various locations, e. g., the textile, plastic or automobile manufacturer or vehicle distribution point. The inspectors take color measurement readings from articles of a particular color. These readings are used to develop color solutions, i. e., different formulations for the same color.
[0005] There are several disadvantages to the present method of color matching. Conventional color laboratories that use human analysis to determine color matching require significant numbers of people, equipment
and materials for identifying pigments and locating a close match from a database. In some cases, an existing formula may provide a close match. In other cases, the formula must be adjusted, mixed, applied and compared to a standard. These steps are repeated until a suitably close match is found. In other cases, no match is found and a formula must be developed from scratch.
Correction of the formula requires a highly skilled technician proficient in the interaction of light with several different pigments.
[0006] Moreover, traditional ! computer software that assists a technician has several disadvantages. Traditional computer software has not proven to be very effective on colors containing"effect pigments."This software is typically based on a physical model of the interaction between illuminating light and the colorant or coating. These models involve complex physics and do not account for all aspects of the phenomena. A traditional approach is to use a model based on the work of Kubleka-Munk or modifications thereof. The model is difficult to employ with data obtained from multi-angle color measuring devices. One particular difficulty is handling specular reflection that occurs near the gloss angle. Another deficiency of the Kubleka-Munk based models is that only binary or ternary pigment mixtures are used to obtain the constants pf the model. Thus, the model may not properly account for the complexities of the multiple interactions prevalent in many paint and colorant recipes.
[0007] The present invention is directed to solving one or more of the problems identified above.
[0008] SUMMARY OF THE INVENTION AND ADVANTAGES A color formulation is comprised of a collection of ingredients and the corresponding amount. Included in the ingredient list are pigment dispersions that impart the color characteristics to the formulation. Color samples having various color formulations over different colorant or paint lines may be converted from one colorant or paint line to another according to the composition of the original formula. Historically, these conversions have been performed by determining conversion factors which when applied to an ingredient amount in one paint or colorant system gives the amount of a corresponding ingredient in the secondary system. These conversion factors where determined by calculating the ratio of the pigment concentration in the secondary system ingredient to the pigment concentration in the primary system ingredient.
[0009] In one aspect of the present invention, a system for converting a first color formula using an artificial intelligence conversion model, is provided.
The model is embodied in a neural network and, in particular, a feed-forward back propagation neural network. The first color formulation is expressed as list of ingredients and amounts in a first system. The neural network is trained using this list for each formulation and corresponding formulations in the second system. The neural network includes an input layer having nodes for receiving input data related to color recipes. Weighted connections connect to the nodes of the input layer and have coefficients for weighting the input data.
An output layer having nodes is either directly or indirectly connected to the weighted connections. The output layer generates output data that is related
to the color formulation in the secondary system. The data of the input layer and the data from the output layer are interrelated through the neural network's nonlinear relationship.
[0010] Neural networks have several advantages over conventional logic- based expert systems or computational schemes. Neural networks are adaptive and provide parallel computing. Further, because neural responses are non-linear, a neural network is a non-linear device, which is critical when applied to non-linear problems. Moreover, systems incorporating neural networks are fault tolerant because the information is distributed throughout the network. Thus, system performance is not catastrophically impaired if a processor experiences a fault.
BRIEF DESCRIPTION OF THE DRAWINGS [0011] Other advantages of the present invention will be readily appreciated as the same becomes better understood by reference to the following detailed description when considered in connection with the accompanying drawings wherein: [0012] Fig. 1 is a block diagram of a system for converting a color formula having an artificial intelligence model, according to an embodiment of the present invention; [0013] Fig. 2 is a diagram depicting a neural network for use in the artificial intelligence model of Fig. 1, according to an embodiment of the present invention; and
[0014] Fig. 3 is a block diagram depicting the training of the color formula conversion neural network of Fig. 2, according to an embodiment of the present invention.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT [0015] Referring to the Figs., wherein like numerals indicate like or corresponding parts throughout the several views, a system 100 and method 300 for converting a color formula, such as paint, pigments, or dye formulations is provided.
[0016] For example, where a paint formulation is used in the repair of an automobile body panel, the input to the system is the formulation in the primary paint system.
[0017] With specific reference to Fig. 1, the system 100 includes an input device 102 for entering a first color formula. Preferably, the first color formula is composed of a plurality of ingredients.
[0018] Preferably, the system 100 is embodied in a computer program run on a general purpose computer (not shown). The input device 102 may be embodied in a user interface for inputting the formulation, such as a keyboard, mouse and/or graphical user interface. Furthermore, the input device 102 may be embodied in an element of a computer system so as to receive the formulation as input from another element of the computer system, such as a computer database, an electronic mail file or other suitable element of the computer system (see below).
[0019] The system 100 of the present invention further includes an artificial intelligence conversion model 104 coupled to the input device 102. The conversion model 104 produces an output signal 106 for communicating a second color formula. The artificial intelligence conversion model 104 may be embodied in a neural network. More specifically, the conversion model 104 may be a back propagation neural network or any other suitable neural network. The output signal 106 may be embodied in a second color formula format, a predicted color formula format or any other suitable format.
[0020] Referring to Fig. 2, an artificial neural network is generally shown at 200. Artificial neural networks 200 are computing systems that model vertebrate brain structure and processes. Artificial neural network techniques are a member of a group of methods that fall under the umbrella of artificial intelligence. Artificial intelligence is commonly associated with logic rule-based expert systems where the rule hierarchies used are reasoned from human knowledge. In contrast, artificial neural networks 200 are self-trained based on experience acquired through data compilation and computation. Thus, artificial intelligence utilizing neural networks 200 is particularly useful in conjunction with complex systems or phenomena where the analysis is complicated, and deriving a model from human knowledge for use in a conventional expert system is a daunting task.
[0021] Although neural networks differ in geometry, activation function and training mechanics, they are typically organized into at least three layers. The first layer is an input layer 220 having one or more input nodes 224,226,228.
The second layer is an output layer 260 having one or more output nodes 264,
266,268. Each output node 264, 266,268 corresponds with an input node 224,226,228. Between the inner and outer layers, there are one or more hidden layers 240, each having one or more hidden nodes 244,246,248 corresponding to an input node and output node pair 224,264,226,266,228, 268. Each input variable is associated with an input node 224,226,228 and each output variable is associated with an output node 264,266,268. Within the neural network 200, data flows in only one direction, such that each node 224,226,228,244,246,266,268 only sends a signal to one or more nodes and receives no feedback.
[0022] The enabling power of a neural network 200 is its connectivity, or the connections between the various nodes 224, 226,228,244,246,266,268. (A configuration technique modeled after the structure of the human brain.) Moreover, because the network is structured, or connected, in such a way as to provide parallel processing (where each node 224, 226, 228,244,246, 266, 268 has connections with other nodes 224,226,228,244,246,266, 268), it is extremely efficient at acquiring and storing experiential knowledge and, then recalling and using that knowledge. More specifically, a node 224, 226,228,244,246,266,268 receives input values, processes them and provides an output. The processing step includes summing the inputs, adding a bias value and submitting this total input to an activation function which limits the magnitude of the output. The connections between the various nodes 224, 226, 228,244,246,266,268 are weighted. An output sent from one node 224,226,228,244,246,266,268 to another is multiplied by the weighting factor associated between those two particular nodes 224,226,228,244,246,
266, 268. The weighting factor represents the knowledge of the system. The system continues to accumulate knowledge and adjust the weighting factor in accordance with training and the further acquisition of knowledge by the network 200. Consequently, the output of the network 200 agrees with the experience of the network 200.
[0023] The neural network 104 of the subject invention is self-trained using formulations in one paint system and the corresponding historical formulations in a secondary paint system. With reference to Fig. 3, a method of training the neural network 104 is illustrated. There are two different types of training (learning) for a neural network 104. In supervised training (or external training), the network 104 is taught to match its output to external targets using data having input and output pairs. In supervised training, the weighting factors are typically modified using a back-propagation method of learning where the output error is propagated back through the network 104. In unsupervised training (or internal training), the input objects are mapped to an output space according to an internal criteria.
[0024] The preferred embodiment of the subject invention neural network 104 is a back propagation neural network 104. In a first process block 302, the first set color formula is input into the neural network 104. In a second process block 304, the neural network 104 converts the first color formula into a second color formula. In a third process block 306, the neural network 104 produces an output signal 106. The output signal 106 is a predicted color formula. In process block 308, an actual color formula is input. In the process block 310, the actual color formula is input and compared to the output signal
106 of the neural network 104, wherein the difference is an error calculation.
In a decision block 312, if the error calculation is sufficiently small, no further action is taken in process block 316. The size of the acceptable error calculation may be determined by experience or be related to an estimation of the uncertainty associated with the formulations. For example, if the uncertainty in the ingredient amount in each formula is Y, and there are on average ten ingredients per formula and the training set consists of 1, 000 formula pairs, then the total or maximum error could be expressed as of Yx10x1, 000. However, where the error calculation is not sufficiently small, the plurality of weighted factors are adjusted based on the output signal 106 in process block 314.
[0025] In another aspect of the present invention, in process block 318, there is a processing of the training data set prior to training of the network. The processing is used to remove inconsistencies or errors in the formulations. An example of an inconsistency is a situation where the historical color formulation in the secondary paint system contains ingredients that would not typically be used in the formulation based on an analysis of the corresponding formulation in the primary paint system. Such inconsistencies may be a result of variability or errors in the color matching process. The preprocessing of the training set may be performed by manual inspection, computational/statistical analysis or via artificial intelligence based techniques.