Chapter 959 Target Contour Recognition

If Chang Haonan says that there is no problem in solving partial differential equations, then at least in China, it is unlikely that someone will easily jump out to question this conclusion.

Therefore, everyone basically agreed that their next step was to find the correspondence between image processing methods and specific partial differential equations.

However, this is still only the first step towards a true ocean surveillance system...or for the entire image recognition and processing technology.

Even the first step in the first action -

Even in the "image segmentation" link, there are still other needs that need to be addressed.

Of course, the ideas proposed by Chang Haonan are indeed novel enough.

Soon, some people began to draw inferences from this:

"Mr. Chang, if a single level set function can divide an image into a background region and a target region, that is, perform a two-phase segmentation, then if we use multiple level set functions at the same time, can we complete the segmentation of a multi-phase image?"

"In theory, of course."

Chang Haonan replied:

"When an image contains more information and cannot be simply divided into two parts: the target and the background, it is necessary to consider the multi-phase variational level set problem."

“However, multiphase problems are a step up in complexity.”

After saying this, he turned around and drew a square on the blackboard.

"We can simply think of each level set equation as defining a region in the image, which is this square."

"When there is only one square, the image is divided into two parts, the inner and the outer. There is no second way of dividing it."

"but."

Chang Haonan drew a second square on the blackboard and marked 1 and 2 in the two figures respectively:

"When the second square appears, there is not one more variable, but two more variables: the second square itself, and the positional relationship between the two squares."

"If the two squares do not overlap, the entire image will be divided into three areas, namely inside No. 1, inside No. 2, and outside No. 1 and No. 2."

“而如果两个正方形重迭，那么整个图像则会被分为四个区域，1号内2号外、1号外2号内、1号内2号内和1号外2号外……”

"In fact, when we set multiple level set functions to divide an image, the image can be divided into 2^N regions at most, and N regions at least. Therefore, the difficulty of classification does not increase linearly, but increases exponentially..."

"..."

This part is actually quite complicated if described in mathematical language, but Chang Haonan explained it completely in plain language, which makes it easy to understand.

"so……"

The person who asked the question just now also understood:

"We also need to specify a level set design scheme first before we can specifically complete the regional division of the multiphase diagram?"

"Yes, but it's not complete either."

Chang Haonan threw the last bit of chalk into the chalk box next to him:

"It seems that there is only one more degree of freedom in the software design logic, but in fact, except for the extreme cases of dividing into N or 2^N regions, additional constraints must be added, and the amount of calculation is quite daunting."

"Even for the simplest multiphase image segmentation, which is to use two level set functions to divide four regions, four local grayscale fitting functions will appear, and then the variational level set method will be used to minimize the total energy functional. This process requires solving eight Euler-Lagrange equations. You can figure out more complicated situations yourself..."

"In addition, the probability of capturing two high-value specific targets in one photo is already very low. Even if this happens by chance, as long as we can identify any one of them and correctly classify the image as the key point and send it back to the ground, the effect will be the same. After all, we still have ground analysts and military commanders who don't just rely on information provided by satellites to fight."

"..."

While Chang Haonan was talking, he simply wrote on the blackboard the solution results of the fitting function and the average gray value in the classic case.

Obviously, unless the circumstances are very coincidental, such a process will not be suitable for completion on a satellite-borne computer, at least within the next few years.

"So, this aspect is not our research focus for the time being?"

Lin Senming put down the pen in his hand, moved his wrist slightly and asked.

"Not necessarily."

Chang Haonan wiped his hands and sat back at the head of the conference table:

"If the equation is relatively stable, then we can actually use the semi-implicit difference scheme to solve it numerically. There is actually no other obstacle except that it takes a very long time..."

"In short, I have ready-made theoretical ideas and computing tools. If you have enough funds, computing power and manpower, Mr. Lin, I can ask Professor Yao Mengna from our laboratory to be in charge of the specific connection with you."

"Although the satellite cannot do this, there is still some value in processing the returned images on the ground."

"In addition, we can broaden our horizons. This type of technology does not necessarily have to be used on satellites. Aircraft, ships, and even vehicle platforms can support computing power far exceeding that of satellites. They can all benefit from it in the future. This does not even include the potential in the civilian field..."

Although Lin Senming was one of Chang Haonan's deputies in the Ocean No. 1 satellite project, the former was a member of the Aerospace Science and Technology Group after all, and there was actually no organizational relationship between the two parties.

Therefore, Chang Haonan could only use the method of drawing pie in the sky to induce people.

The main reason is that his own energy and the resources at hand are always limited.

It would be better to take advantage of this opportunity to give full play to the abilities of the non-staff deputy director of the Science and Technology Commission and mobilize the subjective initiative of the units at the same level.

In this way, when we develop strategic reconnaissance aircraft or other similar needs arise in the future, we will have a ready-to-use product on the shelf.

It can greatly shorten the R&D cycle.

"Well...I'll report this to my superiors when I get back."

Since it has no direct relationship with Ocean No. 1, it is impossible for Lin Senming to admit it immediately.

But it was not difficult to guess from his almost glowing eyes that he was actually quite interested in it.

As for the statement of "reporting to superiors".

It's just a classic delaying tactic.

"Comrades, if you have no further questions."

Chang Haonan said as he turned on his laptop and connected it to the projector next to him:

"Then, I will introduce the mathematical principles of the variational level set model in detail from a theoretical perspective..."

……

There are basically no contestants sitting in this conference room who are completely clueless about the field of image recognition processing.

Therefore, Chang Haonan can skip the introduction of several common image segmentation models to a certain extent.

However, he was unable to finish all the content he had planned in advance on that day.

The two main contents are curve evolution theory and variational level set method, which are indeed a bit difficult to understand.

Especially for some technical personnel with an engineering background, it is extremely abstract.

Fortunately, Lin Senming, who was of great importance, majored in applied mathematics and was able to follow Chang Haonan's ideas throughout the process.

Even so, this meeting... or rather a series of meetings, lasted for four full days and was finally concluded on Friday.

And during this week, another thing closely related to Chang Haonan was happening simultaneously without his knowledge...