Hierarchical Multi-scale Visual Attention Module for Wall Putty Bulge Terminals Detection

Autonomous construction robots intend to accomplish indoor and outdoor construction and decoration operations automatically and efficiently. Among these operations is the wall putty scraping process, which is very time-consuming and tedious when accomplished by humans. Through the putty scraping operation, the putty bulges spread anywhere inside the working frame. To assist autonomous construction robots in finalizing the putty scraping operation successfully, the robots must automatically identify the putty terminal points location. We propose an end-to-end, trainable hierarchical deep learning model based on multi-scale attention feature fusion to localize the putty terminal points precisely. Our model combines the benefits of visual attention and the inception module to boost the model’s feature extraction capabilities. Experimental results showed the effectiveness and performance boost of our proposed model when applied to different ResNet architectures.