Surely at some instance it will give better detection. Because Impulse noise is having a peak noise strength at some instance only. Hence if we simulate at lower noise strength instance, it gives better detection.

I have found that Sometimes it is possible with AWGN and the process is called Stochastic Resonance. But my signal is neither weak nor I had problem with AWGN because by increasing awgn strength my signal detection probability deteriorate and so is normal, but the abnormal thing is with Impulsive noise because as i increase Impulsive noise strength then it increases signal detection probability. And furthermore Impulsive noise doesn't show stochastic resonance property.

@Prakasham

Yes it has only few possibility. But in my case it is happening every-time I run simulation with different parameters. Still I am not able to find out where i am making mistake.

Noise can significantly reduce quantization error. Can this be your case? 

The reason for this (increase Impulsive noise strength then it increases signal detection probability) is that the impulsive noise is wrongly detected as the signal. Moreover, the impulsive noise may be of the featture which the signal also has.

This can increase the probability of detection (Pd), but also increases the probability of false alarm (Pf). It’s better to plot ROC (Receiver Operating Characteristics) curve (Pd vs Pf) for your system (simulation). In comparing variety of systems, we would like to select the system (or develop one) that maximize Pd for every given Pf.

yes it's. the situation is to be able to separate it (the noise) after.

example we have on umts, after gain processing, you could get the useful symbols :-) we use every day.

Since  feature is a high frequency detail, detector wrongly detects high frequency impulsive noise as a feature. To reduce this false detection rate, optimal thresholding or some other tech. could have been used.  But this is not the case for AWGN which is constant through out the spectrum.

Yes, it is possible for noise+signal to be above threshold and signal alone to be below threshold. Noise is not always your enemy. Noise can also attenuate the appearance of irrelevant sharp edges, making an underlying signal relatively more visible. In that case blurring works better, but noise can still work. You appear to be near a fun research area, stochastic algorithms. Lots of wonderful counterintuitive stuff in there.

It is hard to answer your question, since you haven't tell anything about how you implemented your feature detector and how you evaluated performance. It can also be a bug in your program.

It is also known that adding noise (dithering) can improve performance of quantization and in some cases performance of adaptive algorithms, when averaging over many samples and the noise has distribution which is not with more than one peak.

Adding noise to remove non-linear degradation has been carried out in these work:

1. Removing bleed-through degradation:

https://www.researchgate.net/publication/268221733_Degraded_Document_Images_Enhancement_and_Reconstruction_Based_on_Non-local_Sparse_Representation

2. Deblurring:

http://faculty.ucmerced.edu/mhyang/papers/cvpr14_deblur_text.pdf

Conference Paper Degraded Document Images Enhancement and Reconstruction Base...

Dr. Stephen Kay wrote an IEEE paper regarding that very topic:

http://www.ele.uri.edu/faculty/kay/New%20web/downloadable%20files/00809511.pdf

By knowing a little about the noise you're adding, you can improve detector capability, but just by a little bit.

Interesting to have counter intuitive results that can lead to a breakthrough. Lol. But may be we need sometimes to slow down before jumping to conclusions. In fact this is what you are doing now by raising this question. As said above @Tomazic, it can be a bug in the program!. Lost discovery!. Unfortunately. Nevertheless, you have achieved something. You discovered a bug in our way of thinking simulation. Again, we will all have to speculate a lot as we don't have more details. I think some commented above (@Prakasam and @Hiuyue) that this can be the case (sometimes) when your signal and noise (the spikes) are combining forces (in one direction)to overcome the threshold. But the question, if that happened in a number of occasions (randomly) it should be that the reverse (spikes and signal oppose each other leading to error) happens on a number of other occasions (again randomly) so that on the average we have a zero impact. Hence, in this case, I suggest to test your source for randomness and make sure that it is a true random generator not favoring one bit over the other. If passed, as expected (as you are probably using some general random generator in the language you use), make sure that in further stages-e.g. signal shaping and channel parameters of attenuation and -white- noise don't change this randomness)...Also, make sure that your simulation is long enough to obtain steady state results.

On the other hand, we are all assuming that you test signal transmission over a noisy channel (white noise or white noise+spikes) and determining Pe immediately at the output of the threshold detector; i.e. not after some post processing (e.g. error detecting codes) where post processing (sometimes combined with signal coding at the Tx) can lead to optimize detection with shot noise. Shot noise can be easily detected and corrected with the proper coding (e.g. having a long Block code capable of detecting/correcting burst errors resulting from shot noise of certain duration). I hope this helps. Keep us informed. Thanks. @AlDmour.

Two new theorems show how deliberately adding quantizer noise can improve statistical signal detection in array-based nonlinear correlation detection even in case of infinite-variance alpha-stable channel noise.The first theorem gives a necessary and sufficient condition for such quantizer noise to increase the detection probability for a fixed false-alarm probability.The second theorem shows that the array must contain more than one quantizer for a stochastic resonance noise benefit and that the noise benefit improves in the small quantizer noise limit as the number of array quqntizers increases. It further shows that symmetric uniform quantizer noise gives the optimal noise benefit among all symmetric scale-family noise types. 

Addition of noise in the received signal increases the probability of signal detection at the receiver but at the cost of increased BER.

Dear Prof. Kishan. Although Imtiyaz in his question above is talking about Impulsive+AWGN and not about more quantization noise resulting in improved detection probability it is impressive to find similar case of increased noise (of any other type in addition to the AWGN) results in improved detection. Hence, can you please give us references to this kind of research.This is though still looks counter-intuitive, to me at least, especially that in the case of quantization noise which is usually treated as uniform pdf. Thanks. @AlDmour