how to improve performance when running tfjs in web worker?

[libofei2004]

System information

OS Platform and Distribution (e.g., Linux Ubuntu 16.04): MIUI13.0.5 android12
TensorFlow.js installed from (npm or script link): Script Links
TensorFlow.js version (use command below): 4.9.0
Browser version: webview 96

I run tfjs to do object detecting in webview on android mobile phone. For improving the performance , I usually use 6 workers to do object detecting job, but on usual the more workers are executed, the worse performance for one worker I met.

If I use only 1 worker to detect, the waste time is about 60ms - 100ms. If I use 6 worker to detect, the waste time in 1 worker is about 200-300ms in most cases, occasionally the waste time in 1 worker may be about 60ms - 100ms. I don't know how to improve to obtain a stable high performance like 60ms - 100ms when I use 6 workers?
my main code:
index.js

	     ...
             var canvasElement = new OffscreenCanvas(clientWidth, clientHeight);
             function detectImage() {
                         ...                       
			let ctx2 = canvasElement.getContext("2d");
			ctx2.drawImage(V, 0, 0, clientWidth, clientHeight);
                       var img = canvasElement.transferToImageBitmap();
			pool.push(img);
                        setTimeout(function () {
					detectImage();
			}, 40);
              }

              function doDetect(){
                        ...
                        let workerShell = getWorkerPool(); 			
			if(workerShell){
                            let img = pool.shift();
			   doWork(workerShell, img, dealResult);
                        }
              }
            
              function initWorkerPool(num){
		     workerPool = [];
		     for(let i=0;i<num;i++){
			     let worker = new Worker("work.js");
			     workerPool.push({
				    num:i+1,
				    worker:worker,
					busy:false
				 });	
				 worker.postMessage({
						action: "model"
				 });				 
		   }
		}

               function doWork(workerShell, imgdata, callback){
		        let worker = workerShell.worker;
			workerShell.busy = true;		
			worker.onmessage = function(e) {
				let result = e.data.result;			 
				callback(result);
				workerShell.busy = false;
			};		
         	
			worker.postMessage({
					image: imgdata,
					action: "image"
			}, [imgdata]);
		}
		
               function animate() {
			requestAnimationFrame( animate );
			doDetect();
		}
               //simplify code , actually I will wait for all models are loaded in workers and click a button to run detectImage()
                ...
		initWorkerPool(6);
                detectImage();
		animate();

worker.js

importScripts("tf.min4.9.js");
//customvision-tfjs
importScripts("index.umd.js");

var model = new cvstfjs.ObjectDetectionModel();

async function loadModel(){
    await model.loadModelAsync('TensorFlowJs_g_It21/model.json');   
}

self.addEventListener('message', async function (e) {
    
  var action = e.data.action;
  if(action=="image"){
	 var imagedata = e.data.image;
	 var start =  Date.now();
         var result = await model.executeAsync(imagedata);
         var end =  Date.now();
        var time = end-start;
// 60ms - 100ms in only 1 worker, 200-300ms in 6 workers.
        console.log("time",time);
        self.postMessage({result:result, time:time});
  }else if(action=="model"){
	 await loadModel();
	 var result = {state:1};
        self.postMessage(result);
  }
}

I use https://github.xiaoc.cn/microsoft/customvision-tfjs to loadmodel and detect object.