Abstarct: Seismic noise can degrade the continuity of the events on seismic sections and affect the performance of processing steps such as deconvolution, velocity analysis and migration. Among various types of noise in seismic data, random noise is the most important one which covers the main events in a wide range of frequency and time. Therefore, it is difficult to attenuate them by conventional methods. Nonstationary properties of seismic signals make the random noise attenuation more challenging. So far many methods have been introduced for filtering of random noise. Time- frequency peak filtering (TFPF) method has some good benefits because of taking into account the nonstationarity of seismic signals. In this method first the noisy signal is encoded to an analytic signal, and then denoised signal is achieved by estimating instantaneous frequency (IF) of the encoded signal by taking the peak frequency of the time-frequency distribution of the encoded signal. In this method the first choice for computing time-frequency distribution is Wigner-Ville distribution, but there is a problem. As long as IF of the encoded signal is linear, a unit impulse function (δ) is located at the place of IF on time-frequency distribution. Therefore estimated IF is almost close to the real IF. However, if the IF is of higher orders, some other functions with arbitrary shapes are generated at the location of the IF on time-frequency distribution. Therefore estimated IF is far from the real one. In this study, Pseudo Wigner-Ville Distribution (PWVD) was used to compute the time-frequency distribution. Using PWVD makes it possible to select the window length such that IF to be linear within the window. The window length of the PWVD plays a key role in efficiency of the method. When a short length window is chosen, the TFPF method dose not attenuates much noise from seismic data. To attenuate more amount of noise, a longer window length must be chosen. In other words, the longer the window length, the more noise attenuation is possible. Although choosing long window length attenuates more random noise from data but causes some parts of the main events to be attenuated too. More noise attenuation and preserving main events simultaneously needs to make the noisy signal as linear as possible within the window. This goal obviously is met if the frequency of the main events decreases in some ways. Needless to say that frequency of random additive noise must be kept untouched or is increased as frequency of the main events are decreased. Applying dip radial trace transform on seismic data decreases the frequency of main events. Large number of radial traces is needed to prevent aliasing and some distortions are produced because of interpolation. In this study, a modified form of dip radial trace transform was introduced to overcome these problems. The modified method is carried out faster than conventional one and also needs much less interpolation. Firstly, the modified dip radial trace transform is applied on noisy input seismic data. Then the denoised data is obtained by applying the TFPF method on transformed data. A long length window can be used in this method to attenuate much more noise and preserve main events simultaneously. The modified and conventional methods were applied on both synthetic and real data. Comparison of the obtained results shows the modified method is more efficient than conventional one. Moreover applying the method on seismic data with ground roll showed that the method has a good performance in attenuation of coherent seismic noise.