o0o0o0o0o0o0o / image-processing-from-scratch

This project contains some interesting image processing algorithms that were wrote in python and c++ from scratch.

image processing from scratch

what is this?

This repository contains many interesting image processing algorithms that are written from scratch. Read these codes will allow you to have a comprehensive understanding of the principles of these algorithms.

Implementation
All codes were wrote in python3.7 or c++
moudles you may need:
python:

• numpy for matix calculation
• matplotlib for reading and showing images
• opencv2 for some image operations

c++:

• opencv2

Usage
you can always run a python script just by
python script.py

for c++, you need to compile first
cd build
cmake ..
make

when it's done, you are ready to run the executable file by
./program_name parameters
Just make sure you have the images in the right path, and you might wanna modify the code a bit to process another image.
Have fun!

Contents

• canny edge detection
  It is an algorithm that extracts edges of an image.

• hough transform
  It is an algorithm that can theoratically detects shapes that you can write formulas for it.

• harris corner detection
  This algorithm detects corners.

• fast fourier transform
  2-D fourier transform for images using fft.

• sift
  Scale-invariant feature transform, a well-known technique to extract feature points for image matching. Now added c++ version along with SURF and ORB.

• KNN
  Using balanced K-D tree to find k nearest neighbors of K-dimension points.

• PCA&SVD
  Do PCA and SVD using jacobi rotation.(which is accurate but slow)

• Ransac
  Stitch different images together after knowing the sift keypoint pairs.

• watershed
  watershed segmentation algorithm.

• meanshift
  meanshift segmentation algorithm.

• generalized hough transform
  template match of images, detects a given template in an query image. The vote space is implemented with a sparse vector to support big images.

• closed-form image matting
  a classic image matting algorithm proposed in A Closed-Form Solution to Natural Image Matting

• haze removal
  Using dark channel prior and fast guided filter proposed in Single Image Haze Removal Using Dark Channel Prior and Fast Guided Filter

• a lot to be continued...

image-processing-from-scratch's Issues

Image size not reduced with PCA

I tried using your codes for PCA&SVD and tested it on 0.jpg (21.9 kB) image provided, I found out that in spite of using PCA, the size of image 0_.jpg (73.1 kB) is not reduced.
Any leads will be helpful

SIFT 代码中有一些疑问，求指教

up你好，b站来的。有一些SIFT代码问题求指教：
1、GuassianKernel( ) 求高斯卷积核的时候为啥不需要归一化咧？
2、LocateKeyPoint（）函数中，threshold = 0.5contrastThreshold/(n255SIFT_FIXPT_SCALE)，计算差分图像中的噪点阈值时候，这个公式与视频中所讲的0.50.4/n不同，是怎么来的啊？
3、LocateKeyPoint（）函数中，计算邻域极值的时候，我注意到这个邻域不是 33，而是34？这是为啥呢？对于i是i-1，但对于j就变成了j+2，不应该是j+1嘛？
eight_neiborhood_prev = img_prev[max(0, i - 1):min(i + 2, img_prev.shape[0]), max(0, j - 1):min(j + 2, img_prev.shape[1])] eight_neiborhood = img[max(0, i - 1):min(i + 2, img.shape[0]), max(0, j - 1):min(j + 2, img.shape[1])] eight_neiborhood_next = img_next[max(0, i - 1):min(i + 2, img_next.shape[0]), max(0, j - 1):min(j + 2, img_next.shape[1])]

ERROR:slice indices must be integers or None or have an __index__ `method`

Traceback (most recent call last):
  File "HoughTransform.py", line 94, in <module>
    lines = lines_detector_hough(edge)
  File "HoughTransform.py", line 45, in lines_detector_hough
    eight_neiborhood = accumulator[max(0, result[0,i] - halfThetaWindowSize + 1):min(result[0,i] + halfThetaWindowSize, accumulator.shape[0]), max(0, result[1,i] - halfDistWindowSize + 1):min(result[1,i] + halfDistWindowSize, accumulator.shape[1])]
TypeError: slice indices must be integers or None or have an __index__ `method`

小姐姐呢？我就问Ｂ站SIFT视频里面的小姐姐为什么没有？

Value 'info.shape' is unsubscriptable

{
"resource": "/D:/Github/clone path/image-processing-from-scratch/sift/SIFT.py",
"owner": "python",
"code": "unsubscriptable-object",
"severity": 8,
"message": "Value 'info.shape' is unsubscriptable",
"source": "pylint",
"startLineNumber": 507,
"startColumn": 26,
"endLineNumber": 507,
"endColumn": 26
}

这个参数是什么含义啊？“SIFT_FIXPT_SCALE”

会飞的吴克

运行sift.py时报错

Traceback (most recent call last):
File "/home/zsc/桌面/image-processing-from-scratch/sift/SIFT.py", line 557, in
drawLines(X1,X2,Y1,Y2,match[0][:,0],result)
File "/home/zsc/桌面/image-processing-from-scratch/sift/SIFT.py", line 503, in drawLines
img = Lines(img,info)
File "/home/zsc/桌面/image-processing-from-scratch/sift/SIFT.py", line 491, in Lines
result[i,j] = color
ValueError: cannot copy sequence with size 3 to array axis with dimension 4

运行Watershed算法报错

Windows上使用VS2019运行分水岭算法代码时，出现如下错误：

找不出问题所做，请求帮助。

some questions in function convlove()

Hi, I have a question about the function convlove()
"val = (filter * pad_mat[j * strides[1]:j * strides[1] + filter_size[0],
k * strides[0]:k * strides[0] + filter_size[1]]).sum()"
why you multiply "j" with stride[i]? In the "for j in range(0, mat_size[0], strides[1]):" , there is a strides[1]. I'm confused about it. Thank you!