ppu-paddle-ocr/processor/recognition.service.js

Blazing-fast and lightweight PaddleOCR library for Node.js and Bun. Perform accurate text detection, recognition, and image deskew with a simple, modern, and type-safe API. Ideal for document processing, data extraction, and computer vision tasks.

export class RecognitionService{options;debugging;session;toolkit;static BLANK_INDEX=0;static UNK_TOKEN="<unk>";static MIN_CROP_WIDTH=8;constructor(session,options={},debugging={}){this.session=session;this.toolkit=CanvasToolkit.getInstance();this.options={...DEFAULT_RECOGNITION_OPTIONS,...options};this.debugging={...DEFAULT_DEBUGGING_OPTIONS,...debugging}}log(message){if(this.debugging.verbose){console.log(`[RecognitionService] ${message}`)}}async run(image,detection,charactersDictionary){this.log("Starting text recognition process");try{let sourceCanvasForCrop=image instanceof Canvas?image:await ImageProcessor.prepareCanvas(image);let validBoxes=this.filterValidBoxes(detection);let results=await this.processBoxesInParallel(sourceCanvasForCrop,validBoxes,charactersDictionary);return this.sortResultsByReadingOrder(results)}catch(error){console.error("Error during text recognition:",error instanceof Error?error.message:String(error));return[]}}filterValidBoxes(boxes){return boxes.map((box,index)=>({box,index})).filter(({box,index})=>this.isValidBox(box,index))}async processBoxesInParallel(sourceCanvas,boxData,charactersDictionary){let cropsDebugPath=this.debugging.debugFolder+"/crops";if(this.debugging.debug){this.toolkit.clearOutput(cropsDebugPath)}let processingTasks=boxData.map(({box,index})=>this.processBox(sourceCanvas,box,index,boxData.length,cropsDebugPath,charactersDictionary));let results=await Promise.all(processingTasks);return results.filter((result)=>result!==null)}async processBox(sourceCanvas,box,index,totalBoxes,debugPath,charactersDictionary){let start=Date.now();try{let cropCanvas=this.cropRegion(sourceCanvas,box);const{text:recognizedText,confidence}=await this.recognizeText(cropCanvas,charactersDictionary);if(this.debugging.debug){await this.saveDebugCrop(cropCanvas,index,debugPath);this.logProcessingDetails(box,index,totalBoxes,recognizedText,start)}return{text:recognizedText,box,confidence}}catch(e){console.error(`Error processing box ${index+1}: ${e.message}`,e.stack);return null}}sortResultsByReadingOrder(results){return[...results].sort((a,b)=>{let boxA=a.box;let boxB=b.box;if(Math.abs(boxA.y-boxB.y)<(boxA.height+boxB.height)/4){return boxA.x-boxB.x}return boxA.y-boxB.y})}isValidBox(box,index){if(box.width<=0||box.height<=0){console.warn(`Skipping invalid box ${index+1}: w=${box.width}, h=${box.height}`);return false}return true}cropRegion(sourceCanvas,box){return this.toolkit.crop({bbox:{x0:box.x,y0:box.y,x1:box.x+box.width,y1:box.y+box.height},canvas:sourceCanvas})}async saveDebugCrop(cropCanvas,index,outputPath){await this.toolkit.saveImage({canvas:cropCanvas,filename:`crop_${String(index).padStart(3,"0")}.png`,path:outputPath})}logProcessingDetails(box,index,totalBoxes,text,startTime){let processingTime=Date.now()-startTime;this.log(`Box ${index+1}/${totalBoxes}: [x:${box.x}, y:${box.y}, w:${box.width}, h:${box.height}]`+`
	 → "${text}" (processed in ${processingTime}ms)
`)}async recognizeText(cropCanvas,charactersDictionary){const{imageTensor,tensorWidth,tensorHeight}=await this.preprocessImage(cropCanvas);let inputTensor=new ort.Tensor("float32",imageTensor,[1,3,tensorHeight,tensorWidth]);let results=await this.runInference(inputTensor);return this.decodeResults(results,charactersDictionary)}async preprocessImage(cropCanvas){let processor=new ImageProcessor(cropCanvas);let targetHeight=this.options.imageHeight;let originalWidth=processor.width;let originalHeight=processor.height;if(originalHeight===0||originalWidth===0){throw new Error(`Crop dimensions are zero: ${originalWidth}x${originalHeight}`)}let aspectRatio=originalWidth/originalHeight;let resizedWidth=Math.max(RecognitionService.MIN_CROP_WIDTH,Math.round(targetHeight*aspectRatio));processor.resize({width:resizedWidth,height:targetHeight});let imageTensor=this.createImageTensor(processor,resizedWidth,targetHeight);processor.destroy();return{imageTensor,tensorWidth:resizedWidth,tensorHeight:targetHeight}}createImageTensor(processor,width,height){let canvas=processor.toCanvas();let ctx=canvas.getContext("2d");let imageData=ctx.getImageData(0,0,width,height);let pixelData=imageData.data;let numChannels=3;let imageTensor=new Float32Array(numChannels*height*width);for(let h=0;h<height;h++){for(let w=0;w<width;w++){let pixelIndex=(h*width+w)*4;let grayValue=pixelData[pixelIndex];let normalizedValue=(grayValue/255-0.5)/0.5;for(let c=0;c<numChannels;c++){let tensorIndex=c*height*width+h*width+w;imageTensor[tensorIndex]=normalizedValue}}}return imageTensor}async runInference(inputTensor){let feeds={x:inputTensor};let results=await this.session.run(feeds);let outputNodeName=Object.keys(results)[0];let outputTensor=results[outputNodeName];if(!outputTensor){throw new Error(`Recognition output tensor '${outputNodeName}' not found. Available keys: ${Object.keys(results)}`)}return outputTensor}decodeResults(outputTensor,charactersDictionary){let outputData=outputTensor.data;let outputShape=outputTensor.dims;let sequenceLength=outputShape[1];let numClasses=outputShape[2];let dict=charactersDictionary||this.options.charactersDictionary;if(numClasses!==dict.length){console.warn(`Warning: Model output classes (${numClasses}) does not match dictionary length (${dict.length})`)}return this.ctcGreedyDecode(outputData,sequenceLength,numClasses,dict)}ctcGreedyDecode(logits,sequenceLength,numClasses,charDict){let decodedText="";let lastCharIndex=-1;let charConfidences=[];for(let t=0;t<sequenceLength;t++){const{value:maxProb,index:predictedClassIndex}=this.findMaxProbabilityClass(logits,t,numClasses);if(predictedClassIndex===RecognitionService.BLANK_INDEX||predictedClassIndex===lastCharIndex){lastCharIndex=predictedClassIndex;continue}if(this.isValidDictionaryIndex(predictedClassIndex,charDict)){this.appendCharacterToText(predictedClassIndex,charDict,(char)=>{decodedText+=char;charConfidences.push(maxProb)})}else{console.warn(`Decoded index ${predictedClassIndex} out of bounds for charDict (length ${charDict.length}) at t=${t}`)}lastCharIndex=predictedClassIndex}let confidence=charConfidences.length>0?charConfidences.reduce((a,b)=>a+b,0)/charConfidences.length:0;return{text:decodedText,confidence}}appendCharacterToText(index,charDict,appendFn){let char=charDict[index];if(index===charDict.length-1){if(char===RecognitionService.UNK_TOKEN){return}else{appendFn(" ");return}}appendFn(char)}findMaxProbabilityClass(logits,timestep,numClasses){let maxProb=-1/0;let maxIndex=0;for(let c=0;c<numClasses;c++){let prob=logits[timestep*numClasses+c];if(prob>maxProb){maxProb=prob;maxIndex=c}}return{value:maxProb,index:maxIndex}}isValidDictionaryIndex(index,charDict){return index>=0&&index<charDict.length}}import*as ort from"onnxruntime-node";import{Canvas,CanvasToolkit,ImageProcessor}from"ppu-ocv";import{DEFAULT_DEBUGGING_OPTIONS,DEFAULT_RECOGNITION_OPTIONS}from"../constants";