Detection of underground objects remains an important task today, particularly when attempting to recover landmines. A Ground Penetrating Radar (GPR) sends short pulses in time domain to detect underground objects by recording the reflected signal and analyzing its properties. GPR traces acquired at different positions in space are combined together to form B-scan images. The presence of objects produce hyperbolic shape fluctuations in B-scan images that depend on the shape, dielectric properties, and depth of the sensed object, and also on the properties of the medium. Dielectric properties of some objects, like plastic objects, create very small fluctuations in GPR traces. Efficient algorithms to analyze recorded patterns need to be developed and improved based on the inherent variations of the overall conditions. This paper compares the results of three algorithms on three different scenarios for detecting underground objects under noise using B-scan images: Histograms of Oriented Gradients (HOG), 3-row Average Subtraction (3RAS) and Min-max normalization. According to the results, HOG and 3RAS algorithms increase Object Detection Ratio (ODR) from 88% to 93% while decreasing the False Alarm Rate (FAR) considerably. The accuracy is also tested for different image sizes. And for certain algorithms, lower resolution images result in higher accuracy.