This paper introduces a cepstral approach for the detection of landmines from acoustic images. This approach is based on transforming the 2D landmine images to 1D signals using a spiral scan to make object pixels as close as possible to each other after the scan. The Mel-frequency cepstral coefficients (MFCCs) and polynomial shape coefficients are extracted from these 1D signals to form a database of features, which can be used to train a neural network. The discrete cosine transform (DCT) and the discrete wavelet transform (DWT) are also investigated in this paper for the possible extraction of features from these transforms of the original images and/or the original images themselves. The detection of landmines can be performed by extracting features from any new image with the same method used in the training phase. These features are tested with the neural network to decide whether a landmine exists or not. Experimental results show that, the proposed cepstral approach with features extracted from the 2D DCT are the most robust and reliable features in the detection process because of its strong energy compaction property.