池化后的图像尺寸发生变化,计算公式如下:
思路:
1、遍历池化后的图像数据。
2、在原图找到滑动窗口下对应的数据值(max、aveal等)。
以平均池化为例,数据类型为opencv 的mat格式:
void ave_pool(int filter_size, int padding, int stride)
{
Mat src = imread("pic/open_image.jpg", IMREAD_COLOR);
static int yu;
int width = src.cols;
int height = src.rows;
int channels = src.channels();
int _width = (float)(width - filter_size) / stride + 1;
int _height = (float)(height - filter_size) / stride + 1;
Mat out = Mat::zeros(Size(_width, _height), CV_8UC3);
printf("w=%d,h=%d,c%d\n", _width, _height, out.channels());
for (int j = 0; j < _height; j++)
{
uchar *ptrlines = out.ptr<uchar>(j);
for (int i = 0; i < _width; i++)
{
for (int k = 0; k < channels; k++)
{
ptrlines[3 * i + k] = get_max(src, filter_size, i * stride, j * stride, k);
}
}
}
imwrite("out.jpg", out);
imshow("answer", out);
waitKey(0);
destroyAllWindows();
}
uchar get_ave(Mat src, int filter_size, int x, int y, int c)
{
float sum;
for (int j = 0; j < filter_size; j++)
{
uchar *ptr = src.ptr<uchar>(y + j);
for (int i = 0; i < filter_size; i++)
{
sum += ptr[(x + i) * 3 + c];
}
}
uchar ave = sum / (filter_size * filter_size);
return ave;
}
下面输入:ave_pool(9,0,2)
总结:以上效果说明pool对图像具有采样功能,通常用于对特征图像进行降维(特征压缩)。