TFLite 演示模型上的静态图像精度低 [英] Low Accuracy with static image on TFLite demo model

问题描述

我正在尝试使用来自

这是正在分类的连续图像流.

我需要在拍摄图片而不是流后对图像进行分类，然后根据结果采取一些行动.下面是我的方法.

1. 创建一个基本的相机应用
2. 拍照并保存到存储区
3. 保存图像的 uri，然后从URI.
4. 这个drawable然后被转换为位图.
5. 位图大小转换为 224 x 224 以匹配移动网络模型
6. 我收到了来自代码实验室的连续流样本的 0.05 和 0.06 范围内的准确度，这在经过训练的花卉类中提供了 0.80 - 0.90 范围内的准确度

下面是我将位图转换为 224 x 224 大小的代码

private static Bitmap getResizedBitmap(Bitmap bm, int newWidth, int newHeight, boolean isNecessaryToKeepOrig) {int width = bm.getWidth();int height = bm.getHeight();float scaleWidth = ((float) newWidth)/width;浮动 scaleHeight = ((float) newHeight)/高度；//为操作创建矩阵矩阵矩阵 = 新矩阵()；//调整位图大小matrix.postScale(scaleWidth, scaleHeight);//重新创建"新位图Bitmap resizedBitmap = Bitmap.createBitmap(bm, 0, 0, width, height, matrix, false);if(!isNecessaryToKeepOrig){bm.recycle();}返回调整大小的位图；}

即使我将原始位图传递给分类器，分类器本身正在将图像转换为 224 x 224，结果仍然相同.我应该对图像进行更多额外处理还是需要更改任何配置在模型中?

解决方案

我认为问题在于平滑概率的 applyFilter().去掉就可以正常显示概率了.

StringclassifyFrame(Bitmap bitmap) {...//平滑结果//应用过滤器();<--删除它...}

I'm trying the TFLite implementation for Image Classification using Mobile Net Transfer Learning example from TensorFlow for Poets 2

I'm able to succesfully complete the transfer learning using the four flower samples in the code lab and got the below screen

This is a continuous stream of images that's being classified.

I need to classify the image after taking the picture instead of stream and then take some action based on the result. Below is my approach for this.

1. Create a basic camera app
2. Take a picture and save it to storage
3. The uri of the image is saved and then a drawable is created from the URI.
4. This drawable is then converted to a bitmap.
5. The bitmap size transformed to 224 x 224 to match the input of the Mobile Net model
6. I'm receiving the accuracy in the ranges of 0.05 and 0.06 against the continuous stream sample from the Code Labs which gives accuracy in the range of 0.80 - 0.90 in the trained flower classes

Below is the code where I transform the bitmap to 224 x 224 size

private static Bitmap getResizedBitmap(Bitmap bm, int newWidth, int newHeight, boolean isNecessaryToKeepOrig) {
    int width = bm.getWidth();
    int height = bm.getHeight();
    float scaleWidth = ((float) newWidth) / width;
    float scaleHeight = ((float) newHeight) / height;
    // CREATE A MATRIX FOR THE MANIPULATION
    Matrix matrix = new Matrix();
    // RESIZE THE BIT MAP
    matrix.postScale(scaleWidth, scaleHeight);

    // "RECREATE" THE NEW BITMAP
    Bitmap resizedBitmap = Bitmap.createBitmap(bm, 0, 0, width, height, matrix, false);
    if(!isNecessaryToKeepOrig){
        bm.recycle();
    }
    return resizedBitmap;
}

The results turn out to be same even when I pass down the original bitmap to classifier which itself is converting the image to 224 x 224. Should I be doing some more additional processing on the images or do I need to change any configuration in the model ?

解决方案

I think the problem is applyFilter() which smooth the probability. Just remove it then the probability should be showed as normal.

String classifyFrame(Bitmap bitmap) {
...
// smooth the results
//applyFilter(); <--remove it

...
}

这篇关于TFLite 演示模型上的静态图像精度低的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！