Abstract:
This research presents a method to detect and recognize buildings in terrestrial images. High-resolution terrestrial images are normally taken from land survey vehicles. These images and other surveyed data along roads are needed by many agencies that require new data as time passes by. Land use in rural area is an example that needs information about buildings and can benefit from terrestrial images. The proposed method was aimed to detect and recognize buildings in terrestrial images to benefit the above needs. The method consists of two stages. The first stage is building detection. It removes unwanted objects, performs image segmentation and finds regions of interest. Image processing techniques such as greenness removal, sky removal, color segmentation, color detection, shape detection are used. After that, building parts detection, projection profiles finding and the building determination are performed. The second stage is building recognition. It begins with image normalization, and uses convolutional neural network to recognize buildings. The method can identify a partial building if the whole building is not shown in an image. The training set contained 3,995 building images and 3,348 images without building from many provinces in Thailand. The proposed method was tested on 3,936 images (1,832 images with buildings and 2,104 images without buildings). The images were from Google Street View. The accuracy was determined by human inspection. The method gave promising results with an average of 87.50% accuracy for images with buildings and an average of 97.60% accuracy for images without building.