Encoder and decoder network with ResNet-50 and global average feature pooling for local change detection

Background subtraction is a prevalent way of dealing with detecting the local changes from video scenes. Background subtraction divides an image frame into foreground and background. The proposed scheme is a unique attempt to detect the local changes in video using a combination of the feature pooling module (FPM) with a ResNet-50 encoder–decoder network. In this context, we proposed a robust encoder–decoder structured deep learning network that is trained with limited training data. The proposed scheme has several folds of novelties including as mentioned below. The use of the feature pooling module with the ResNet-50 encoder–decoder network is the first attempt to use background subtraction in complex video scenes. In the proposed scheme the weights of the ResNet-50 network are learnt by using the transfer learning mechanism. Further, due to the use of a selected number of layers in ResNet-50 architecture with a fewer number of trainable parameters, the proposed architecture becomes less complex as compared to competitive architecture like VGG-16. The proposed ResNet-50 encoder with the FPM module is capable of extracting relevant multi-scale features for local change detection from complex videos. The said encoder uses residual connections between the layers and is hence capable of extracting meaningful multi-scale features with a fewer number of parameters and a higher number of layers. We finally used an up-sampling in the decoder to learn a mapping from the feature space to the image-frame space. The model takes an RGB image frame as the input and generates a foreground segmented probability mask for the corresponding image. To evaluate our model, we have tested it on the three popular benchmark databases. The robustness of the proposed scheme is evaluated by comparing its results with twenty-eight state-of-the-art techniques. The evaluation of the results is carried out using visual and eight quantitative evaluation measures.