Fruit obj=new Orange(); and Orange obj=new Orange(); if both works identically in my code, which is the less coupling one?

The coupling here happens between Orange and the rest of your program. If you retain an instance of Orange, then you could conceivably call method specific to Orange and not of Fruit. This is bad for two reasons: A) using methods specific to Orange means you could not act on that instance simply as a Fruit instance (violates Liskov Substitution Principle), and B) it implies that "Orange" is not a "Fruit" in that it behaves differently.

Using a more concrete example, suppose you need a new list and so you do:

ArrayList<String> list = new ArrayList<String>(); 

You decide you want to pass this to a method, and so of course your method becomes:

public void handleList(ArrayList<String> list) { // ... } 

If you think about it, handleList method doesn't particularly care that it is an instance of ArrayList or even List most likely. It need only deal with every object in that collection at least once. So you could safely rewrite it as the following and it would allow you to potentially pass other implementations (your own included!):

public void handleList(Collection<String> list) { // ... } 

If you wanted to take this a step further, you would use a wildcard (?) in the place of String, and call toString() on every item in the collection. The point is that we're not being obliged to use an instance of ArrayList. We could use LinkedList or HashSet or TreeSet instead.

Returning back to your example, ideally in your program, you should care about the fact that it uses Orange as opposed to Grape or Apple classes exactly in one place in your program. After this point, you should only use these instances as Fruit, or to put it another way, you no longer care how it is implemented, you only care how it is used.

That said, it can be convenient to have an abstract common class with common functionality between implementations. This is not a substitute for an interface. An interface defines how it will be used. An abstract class is meant to reduce and simplify your implementations. Conceivably you could have multiple abstract classes with themselves each with several concrete implementations, all of which deriving from a single interface that you use throughout your program.

For more reading, take a look at this article explaining a bit about Liskov Substitution Principle, specifically the part regarding the Open closed principle. You should have a clear understanding in your program where you're defining objects and where you're using them. If they're not in the same location, you should most definitely be using interfaces to define how they're used.

As a general principle, you should use the smallest amount of information necessary.

In your example, if Orange implements the interface Fruit, and if your code works fine without the knowledge that the object is an Orange object, then it should be declared as a Fruit interface. This means there is no need to change anything if you want to have a Strawberry object, except changing the first line to "Fruit obj = new Strawberry();". Of course if parts of the following code expect an Orange object and use methods that are not present in the Fruit interface, then you start with "Orange obj = new Orange();" And you know that your code won't just work with a Strawberry object. That's tough, it requires code changes, but these code changes are necessary.

If you change the interface that Orange implements - well, you can't just change that interface. You will implement both interfaces. Or you will remove the methods that are needed for Fruit and add methods for the new interface. Your code creating the Orange object isn't going to work anymore. Declaring Orange obj = new Orange(); wouldn't have helped you because most likely you were calling interface methods that are not there anymore. Changing the interface is a major change with major consequences. Your FruitSalad class is unlikely to be able to work with an Orange object if it doesn't implement the Fruit interface.

This comes down to the question of

Why do we abstract anyways?

The point of abstraction is to do more with less. When you see code as rules like don't talk to men you don't know who wear dark suits and black hats and don't talk to strangers you have two rules, the latter being more abstract than the former.

So you end up with an abstract rule having less words, being more general and so dealing with more cases.

But what about talking to doctors which may be a) strangers and b) men?

Then the rule doesn't make sense and you have to deal with exceptions to the rule.

What has this to do with your problem?

Having an interface of Fruit you made an abstraction. And every instruction dealing with $things which behave like fruits is very general and could be applied to many fruity things.

You do not want to write code for every single Fruit you know of.

But as the example with above, this falls apart when you need Orange behavior, which may be richer than only fruity behavior.

So "implement to an interface" is a rule of thumb which means:

It is a good idea to group objects sharing a common behavior together with an interface and write instructions for these objects to write less code than for each individual object. But sometimes this rule doesn't hold.

P.S.: "Coupling" means in this "depending on specialization" or being less general. So when you make general rules, you avoid specialization = coupling.