Algorithm of Contour Detection Improvement

Contour Detection is a very good method. It detects the contours in the picture and sorts the by the differences of local color. But because it's sensitivy, it may lead error. Below is a result of the bottle.

So, we can see, The method can cut the contour of the objects. We can see they are bottles. But we do not want these results. We want a full bottle rather than a broken bottle.

To solve this problem, we must combine the objects of a bottle. For example, To right bottle, we want to combine its neck and body. But how could we do this?

Obviously, we can see that the neck and body is continusous. So we can use this nature to combine them. But unfortunately, what we want to combine is not every connected pair. 

So let's see the following pictures:

Of course we want to combine the former rather than the latter. But what is the differencebetween them? I think, the biggest difference of them is that the former's contour is continuous. But the latter's is not. We can see that the degree in the latter is almost vertical. So, we can use this natrue to detect them.

Moreover,  we can see the former's intersection line is very flat but tht later is not. So, we find another evidence. But this is not very absolute. we may fail in some cases. But is can be a optional choice.

Thirdly, we can also make the judge that if we combine the two objects, the convexity of the former is larger than that of the latter. It is clear. The former is almost flat and the latter is not.

So, use these natrue, we can combine the objects.