Single-pixel imaging system attracts a lot of attentions because of its special imaging method, but its target recognition method in noisy environment has not been studied deeply. Aiming at this problem, the signal sequences obtained by the bucket detector and the corresponding formed two-dimensional images were used as the training samples for deep learning to identify targets in noisy environments. By comparing the recognition results of these two methods, it was found that when the sampling rate was low, the former one could obtain a higher recognition rate even in a strong noise environment; while for the latter one, although the recognition rate was relatively stable, its preprocessing time was high, so the former one was more suitable for target recognition in high-speed imaging. In addition, for the method using only the bucket detector signal as the training samples, the effect of target sparsity on its recognition accuracy was also analyzed. It was found that when the external noise and sampling rate were fixed, the higher the sparsity of the target, the higher the recognition accuracy was. This paper can be used as the reference for the selection of single pixel system recognition methods in noisy environments.