ppu-paddle-ocr/processor/detection.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 DetectionService{options;debugging;session;static NUM_CHANNELS=3;constructor(session,options={},debugging={}){this.session=session;this.options={...DEFAULT_DETECTION_OPTIONS,...options};this.debugging={...DEFAULT_DEBUGGING_OPTIONS,...debugging}}log(message){if(this.debugging.verbose){console.log(`[DetectionService] ${message}`)}}async run(image){this.log("Starting text detection process");try{let canvasToProcess=image instanceof Canvas?image:await ImageProcessor.prepareCanvas(image);if(this.options.autoDeskew){this.log("Auto-deskew enabled. Performing initial pass for angle detection.");let angle=await this.calculateSkewAngle(canvasToProcess);this.log(`Detected skew angle: ${angle.toFixed(2)}°. Rotating image at ${-angle.toFixed(2)}° (to ${-angle>1?"right":"left"})...`);let processor=new ImageProcessor(canvasToProcess);let rotatedCanvas=processor.rotate({angle}).toCanvas();processor.destroy();canvasToProcess=rotatedCanvas;if(this.debugging.debug){await CanvasToolkit.getInstance().saveImage({canvas:canvasToProcess,filename:"deskewed-image-debug",path:this.debugging.debugFolder})}}let input=await this.preprocessDetection(canvasToProcess);let detection=await this.runInference(input.tensor,input.width,input.height);if(!detection){console.error("Text detection failed (output tensor is null)");return[]}let detectedBoxes=this.postprocessDetection(detection,input);if(this.debugging.debug){await this.debugDetectionCanvas(detection,input.width,input.height);await this.debugDetectedBoxes(canvasToProcess,detectedBoxes)}this.log(`Detected ${detectedBoxes.length} text boxes in image`);return detectedBoxes}catch(error){console.error("Error during text detection:",error instanceof Error?error.message:String(error));return[]}}async deskew(image){this.log("Starting image deskewing process");let canvasToProcess=image instanceof Canvas?image:await ImageProcessor.prepareCanvas(image);this.log("Performing initial pass for angle detection.");let angle=await this.calculateSkewAngle(canvasToProcess);this.log(`Detected skew angle: ${angle.toFixed(2)}°. Rotating image at ${-angle.toFixed(2)}° (to ${-angle>1?"right":"left"})...`);let processor=new ImageProcessor(canvasToProcess);let rotatedCanvas=processor.rotate({angle}).toCanvas();processor.destroy();if(this.debugging.debug){await CanvasToolkit.getInstance().saveImage({canvas:rotatedCanvas,filename:"deskewed-image-debug",path:this.debugging.debugFolder})}return rotatedCanvas}async calculateSkewAngle(canvas){let input=await this.preprocessDetection(canvas);let detection=await this.runInference(input.tensor,input.width,input.height);if(!detection){this.log("Skew calculation failed: no detection output from model.");return 0}const{width,height}=input;let probabilityMapCanvas=this.tensorToCanvas(detection,width,height);if(this.debugging.debug){await CanvasToolkit.getInstance().saveImage({canvas:probabilityMapCanvas,filename:"deskew-probability-map-debug.png",path:this.debugging.debugFolder})}let deskewService=new DeskewService(this.options,this.debugging);let result=await deskewService.calculateSkewAngle(probabilityMapCanvas);return result}async preprocessDetection(canvas){const{width:originalWidth,height:originalHeight}=canvas;const{width:resizeW,height:resizeH,ratio:resizeRatio}=this.calculateResizeDimensions(originalWidth,originalHeight);let processor=new ImageProcessor(canvas);let resizedCanvas=processor.resize({width:resizeW,height:resizeH}).toCanvas();processor.destroy();let width=Math.ceil(resizeW/32)*32;let height=Math.ceil(resizeH/32)*32;let paddedCanvas=this.createPaddedCanvas(resizedCanvas,resizeW,resizeH,width,height);let tensor=this.imageToTensor(paddedCanvas,width,height);this.log(`Detection preprocessed: original(${originalWidth}x${originalHeight}), `+`model_input(${width}x${height}), resize_ratio: ${resizeRatio.toFixed(4)}`);return{tensor,width,height,resizeRatio,originalWidth,originalHeight}}calculateResizeDimensions(originalWidth,originalHeight){let MAX_SIDE_LEN=this.options.maxSideLength;let resizeW=originalWidth;let resizeH=originalHeight;let ratio=1;if(Math.max(resizeH,resizeW)>MAX_SIDE_LEN){ratio=MAX_SIDE_LEN/(resizeH>resizeW?resizeH:resizeW);resizeW=Math.round(resizeW*ratio);resizeH=Math.round(resizeH*ratio)}return{width:resizeW,height:resizeH,ratio}}createPaddedCanvas(resizedCanvas,resizeW,resizeH,targetWidth,targetHeight){let paddedCanvas=createCanvas(targetWidth,targetHeight);let paddedCtx=paddedCanvas.getContext("2d");paddedCtx.drawImage(resizedCanvas,0,0,resizeW,resizeH);return paddedCanvas}imageToTensor(canvas,width,height){let ctx=canvas.getContext("2d");let imageData=ctx.getImageData(0,0,width,height);let rgbaData=imageData.data;let tensor=new Float32Array(DetectionService.NUM_CHANNELS*height*width);const{mean,stdDeviation}=this.options;for(let h=0;h<height;h++){for(let w=0;w<width;w++){let rgbaIdx=(h*width+w)*4;let tensorBaseIdx=h*width+w;for(let c=0;c<DetectionService.NUM_CHANNELS;c++){let pixelValue=rgbaData[rgbaIdx+c]/255;let normalizedValue=(pixelValue-mean[c])/stdDeviation[c];tensor[c*height*width+tensorBaseIdx]=normalizedValue}}}return tensor}async runInference(tensor,width,height){let inputTensor;try{this.log("Running detection inference...");inputTensor=new ort.Tensor("float32",tensor,[1,3,height,width]);let feeds={x:inputTensor};let results=await this.session.run(feeds);let outputTensor=results[this.session.outputNames[0]||"sigmoid_0.tmp_0"];this.log("Detection inference complete!");if(!outputTensor){console.error(`Output tensor ${this.session.outputNames[0]}  not found in detection results`);return null}return outputTensor.data}catch(error){console.error("Error during model inference:",error instanceof Error?error.message:String(error));throw error}finally{inputTensor?.dispose()}}tensorToCanvas(tensor,width,height){let canvas=createCanvas(width,height);let ctx=canvas.getContext("2d");let imageData=ctx.createImageData(width,height);let data=imageData.data;for(let y=0;y<height;y++){for(let x=0;x<width;x++){let mapIndex=y*width+x;let probability=tensor[mapIndex]||0;let grayValue=Math.round(probability*255);let pixelIdx=(y*width+x)*4;data[pixelIdx]=grayValue;data[pixelIdx+1]=grayValue;data[pixelIdx+2]=grayValue;data[pixelIdx+3]=255}}ctx.putImageData(imageData,0,0);return canvas}postprocessDetection(detection,input,minBoxAreaOnPadded=this.options.minimumAreaThreshold||20,paddingVertical=this.options.paddingVertical||0.4,paddingHorizontal=this.options.paddingHorizontal||0.6){this.log("Post-processing detection results...");const{width,height,resizeRatio,originalWidth,originalHeight}=input;let canvas=this.tensorToCanvas(detection,width,height);let processor=new ImageProcessor(canvas);processor.grayscale().convert({rtype:cv.CV_8UC1});let contours=new Contours(processor.toMat(),{mode:cv.RETR_LIST,method:cv.CHAIN_APPROX_SIMPLE});let boxes=this.extractBoxesFromContours(contours,width,height,resizeRatio,originalWidth,originalHeight,minBoxAreaOnPadded,paddingVertical,paddingHorizontal);processor.destroy();contours.destroy();this.log(`Found ${boxes.length} potential text boxes`);return boxes}extractBoxesFromContours(contours,width,height,resizeRatio,originalWidth,originalHeight,minBoxArea,paddingVertical,paddingHorizontal){let boxes=[];contours.iterate((contour)=>{let rect=contours.getRect(contour);if(rect.width*rect.height<=minBoxArea){return}let paddedRect=this.applyPaddingToRect(rect,width,height,paddingVertical,paddingHorizontal);let finalBox=this.convertToOriginalCoordinates(paddedRect,resizeRatio,originalWidth,originalHeight);if(finalBox.width>5&&finalBox.height>5){boxes.push(finalBox)}});return boxes}applyPaddingToRect(rect,maxWidth,maxHeight,paddingVertical,paddingHorizontal){let verticalPadding=Math.round(rect.height*paddingVertical);let horizontalPadding=Math.round(rect.height*paddingHorizontal);let x=rect.x-horizontalPadding;let y=rect.y-verticalPadding;let width=rect.width+2*horizontalPadding;let height=rect.height+2*verticalPadding;x=Math.max(0,x);y=Math.max(0,y);let rightEdge=Math.min(maxWidth,rect.x+rect.width+horizontalPadding);let bottomEdge=Math.min(maxHeight,rect.y+rect.height+verticalPadding);width=rightEdge-x;height=bottomEdge-y;return{x,y,width,height}}convertToOriginalCoordinates(rect,resizeRatio,originalWidth,originalHeight){let scaledX=rect.x/resizeRatio;let scaledY=rect.y/resizeRatio;let scaledWidth=rect.width/resizeRatio;let scaledHeight=rect.height/resizeRatio;let x=Math.max(0,Math.round(scaledX));let y=Math.max(0,Math.round(scaledY));let width=Math.min(originalWidth-x,Math.round(scaledWidth));let height=Math.min(originalHeight-y,Math.round(scaledHeight));return{x,y,width,height}}async debugDetectionCanvas(detection,width,height){let canvas=this.tensorToCanvas(detection,width,height);let dir=this.debugging.debugFolder;await CanvasToolkit.getInstance().saveImage({canvas,filename:"detection-debug",path:dir});this.log(`Probability map visualized and saved to: ${dir}`)}async debugDetectedBoxes(image,boxes){let canvas=image instanceof Canvas?image:await ImageProcessor.prepareCanvas(image);let ctx=canvas.getContext("2d");let toolkit=CanvasToolkit.getInstance();for(let box of boxes){const{x,y,width,height}=box;toolkit.drawLine({ctx,x,y,width,height})}let dir=this.debugging.debugFolder;await CanvasToolkit.getInstance().saveImage({canvas,filename:"boxes-debug",path:dir});this.log(`Boxes visualized and saved to: ${dir}`)}}import*as ort from"onnxruntime-node";import{Canvas,CanvasToolkit,Contours,createCanvas,cv,ImageProcessor}from"ppu-ocv";import{DEFAULT_DEBUGGING_OPTIONS,DEFAULT_DETECTION_OPTIONS}from"../constants.js";import{DeskewService}from"./deskew.service.js";