Lack of security is a key risk factor to the government and common people for the past ten years. So far, an efficient security system is not yet available to tackle this situation. Hence, there arises a need for developing a landmine detection system using convolutional neural network (CNN) to offer solution to this real-time crisis. The methodology includes training the CNN with various open-source images of landmines downloaded from the Internet. On field, specialized drones were used for capturing the aerial infrared images of the terrain, where buried landmines are detected during surveillance. Software like free online label maker was used to comment on the images. Extensible markup language (XML) was used to save the results. This XML file was converted to comma-separated value (CSV) files, which are represented in the form of rows and columns. Rows represent each element, while the columns represent the shape and location of the landmines. The shape and location include weight, height, orientation-X, orientation-Y, and intensity values at (Xmin, Ymin) and (Xmax, Ymax) of the landmines. Target values were assumed, created, and mapped to a particular value, either as ‘0’ or ‘1’ for anti-personnel and anti-tank mines. Training was carried out for 300 epochs with a learning rate of 0.00124 for ResNet50, ResNet10, InceptionV1, and the proposed multi-magnification deep residual network (MM-ResNet) model. The trained model was tested to determine the success rate during training and testing process. The proposed model performed better than the other three models when trained and tested with the same datasets.