Introduction
Recently, there are many live cameras/CCTVs (closed-circuit television) has been installed in many places such as offices, roads, and homes, in order to help human work. There are various purposes of installing CCTVs, such as security, surveillance, and data analysis. One of the challenges of installing CCTVs is about data management. Most CCTVs are set to take data/image near realtime conditions. If a CCTV takes data/image every minute and each data size is 50 KB, then we need 72MB of hard disk capacity per day, or 26.28 GB per year. If the purpose of using CCTV is to analyze data, we need to take data in a long span. In this scale we can say this is big data, therefore we need to compress data.
Compressing live camera data are commonly done in the following two phases: In the first phase, the CCTV compress the data which will be delivered into the data centre. This is commonly done by JPEG compression. In the second phase, we compress data in the data centre. What I want to explain in this article is a compression which is done in the second phase. In this phase, images are commonly compressed by image subtraction method, then the run-length enconding method. Here we increase the compression rate of the method above by calculating the similarity of images.
Required Application and Libraries
- Python 3.5
- OpenCV 3.1
- scikit-image (compare_ssim) 0.13.0
- PIL
Tested on macOS Sierra 10.12.2 and Linux (Ubuntu 16.04)
Implementation
1. Explanation about the compression method
We combine two methods; First, the image subtraction method, then the run-length compression method. The image subtraction method increases the number of redundant pixels by subtracting pixels from two similiar images, we then compress the redundant pixels by the run-length encoding method. The formula to subtract the image is below:
C = B - A
which is:
C is subtracted image (image member)
B is subtraction subject
A is subtraction object (image key)
The algorithm of image subtraction uses an iteration that compares the similarity between two images. If the similarity is less than the threshold then the image member will become image key, else, the image key will be the same until the iteration process finds an image member that has similarity bigger than threshold. The next step, the program compress image key and image member using run-length encoding compression.
Bellow is the pseudocode:
image key=""
while image(n) < total image
image member=""
if (keyfile == "")
image key = image(n)
image member = image(n+1)
else
image key = image(n)
endif
if similarity(image key, image member) < thresshold
if (image key == image(n))
compress(image key)
endif
compress (image member)
image key = image member
continue
endif
image subtracted = image member - image key
compress (image subtracted)
if (image key == image(n))
compress(image key)
endif
end
Based on that algorithm, the compression rate depends on the total number of image keys and image members. The more image member and less image key that are created, the higher rate compression that will be got.
2. How does image subtraction work?
The image subtraction method uses the OpenCV subtraction function. If there is a same pixel between two images, then the pixel will be colored black, or (0,0,0) in RGB form. Below is the sample of image subtraction using OpenCV and the result:
img1 = cv.imread("./img1.jpeg")
img2 = cv.imread("./img2.jpeg")
img3 = cv.subtract(img2, img1)
Here we will compare the original image and the subtracted image using OpenCV-im.load():
im = Image.open("./data/img1.jpeg")
pix = im.load()
w,h = im.size
total_ori = 0
for i in range (0,w):
for j in range (0,h):
if "(0, 0, 0)" in pix[i,j]:
total_ori = total_ori + 1
im2 = Image.open("./data/img3.jpeg")
pix = im2.load()
w,h = im2.size
total_subtraction = 0
for i in range (0,w):
for j in range (0,h):
if "(0, 0, 0)" in pix[i,j]:
total_subtraction = total_subtraction + 1
print ("original image 0 pixel: %d"% total_ori)
print ("subtraction image 0 pixel: %d"% total_subtraction)
Result:
original image 0 pixel: 25
subtraction image 0 pixel: 112833
Based on the result above, it is clear that the image subtraction method can increase the number of consecutive pixels of identical color. Then the next step is to compress the subtracted images by run-length encoding compression.
3. How does the run-length encoding work?
Run-length encoding encodes images. There are many types of run-length encoding; here we use Packbits. Below is a sample code:
foo=Image.open(filein)
foo.save(fileout, compression = 'packbits')
How does it actually work?
Let us explain the method using an example.
Here is a 16-byte image:
W W W W
W B B W
W B B W
W W W W
Compression Result:
5W 2B 2W 2B 5W
The 16-byte image above will be written as a file with the following format/sequence: "W W W W W B B W W B B W W W W W". Based on the sequence, it is easy to understand that there are redundant pixels in the sequence. Packbits algorithm encodes the redundant pixels by storing series (run) of identical pixels color. In our case, the sequence "W W W W W B B W W B B W W W W W" will encode as "5W 2B 2W 2B 5W". Thus, the total number of bytes decreases from 16 bytes to 10 bytes.
Result
| image similarity | original size | compression size (KB) | compression rate (KB) |
|---|---|---|---|
| neighbor image | 1542076 | 1183892 | 23.22738957 |
| 90% | 1542076 | 1393184 | 15.53101144 |
| 80% | 1542076 | 1155136 | 25.09214851 |
| 70% | 1542076 | 1037132 | 32.74443024 |
| 60% | 1542076 | 1015576 | 34.14228611 |
| 50% | 1542076 | 993840 | 35.55181457 |
| 45% | 1542076 | 1013148 | 33.52675225 |
| 40% | 1542076 | 1025068 | 34.2997362 |
| 35% | 1542076 | 1120552 | 27.33483953 |
The best compression rate is obtained when the image similarity is 50%
Summary
The key point to get higher compression rate with this method is to increase the number of image members and to decrease the number of image keys. However, we have to maintain the similarity between image keys and image members. The more similar the images, the higher compression rate we will get from Packbits.