Functional near-infrared spectroscopy (fNIRS) presents an affordable and light-weight method to monitor the cerebral hemodynamics of the brain. However, noise and artefacts hamper the analysis of fNIRS signals. Thus, the signal quality assessment is a crucial step when planning fNIRS experiments. Currently no standardized method exists for the evaluation. Commonly used visual inspection of the signals is time consuming and prone to subjective bias. Recently use of machine learning and deep learning approaches have been applied for the fNIRS signal quality assessment, showing promising results. However, currently there are only a few experiments which have investigated the use of these approaches to evaluate fNIRS signal quality. In this human brain study, we utilized previously developed deep learning approach used for the assessment of PPG signal quality with short-time Fourier transform (STFT) to evaluate the quality of raw fNIRS signals with wavelengths 690 nm, 810 nm, 830 nm and 980 nm. The data was collected from 38 subjects with a two-channel fNIRS device, measured during breath hold protocol in sitting position. A total of 10,144 segments were extracted using a window of 10 seconds length without overlap and annotated for SQA by three independent evaluators. The segments were transformed with STFT, and further processed into 2D images. The images were used as input data for CNN deep learning network, and the output further used to classify the segments as acceptable or unacceptable. The results show high potential of using DL approach for fNIRS signal quality assessment with classification accuracy of 87.89 %.