Résumé: Change detection (CD) is rapidly becoming a fundamental approach due to its vast applications in image or video analysis. Siamese convolutional networks (SCN) have recently been used to detect changes. However, these approaches are unable to eliminate drawbacks such as coarse borders and empty holes because of the lack of spatial and other necessary information. So, the main goal of this paper is to develop a network that captures necessary features and has a better feature fusion technique for detecting changes. This is achieved by the proposed Siamese Capsule Multi-level Attention-based CD Network (SCMACDnet). Initially, the Siamese Capsule Encoder-Decoder Network (SCEDnet) is introduced to extract spatial features effectively using capsule attention block (CAB). Afterward, Change Atrous Feature Fusion (CAFF) is introduced to hierarchically combine them by exploring the fusion of channel pairs at several feature levels. Also, multi-scale data from the feature map (FM) is captured with a quicker inference speed using Stride Atrous Convolutional Pooling (SACP) in CAFF. Finally, an effective Reverse Dual Attention (RDA) module is introduced to highlight the changed areas by considering both the channel and pixel-wise correlation (PC) to preserve the shape and edges of the changed areas. Experiments conducted on the CDnet 2014 dataset demonstrate that the SCMACDnet method outperforms other existing methods in terms of precision (.963), sensitivity (.960), the F1 measure (.95), and the Percentage of Wrong Classification (.212). Further, the ablation study validates the efficiency of the proposed SCEDnet, CAFF, and RDA. The results show that the proposed SCMACDnet successfully identifies changes with accurate boundaries and shapes.

Langue: Anglais