Image Inpainting based Compact Hash Code Learning using Modified U-Net

As a result of the widespread use of imaging devices, many images are recorded every second. Access to these recorded images becomes more difficult as the number increases. Various content-based image retrieval (CBIR) methods are suggested to solve this problem, but more is still needed. Unlike retrieval methods, which require intense hardware power and require long periods, the proposed study is very practical. The proposed framework produces hash codes of the desired length using a modified U-net model without the need for any label information. The proposed framework first randomly deletes 30% of the pixels of the images in the data. These pixel-deleted images are used as the input of the modified U-net, and the original images as the output. The purpose of this process is to learn all the features of the image without the need for the label value. In the second phase of the proposed framework, the traditional U-net architecture is modified for inpainting. The modification is done using reconstruction loss and adversarial loss. Finally, the hash code acquisition procedure is applied by taking the features from the modified U-net middle layer. The proposed framework has been tested in both natural and medical images. Gray level and color experiments are carried out with the images in these two data. The results obtained are highly promising and competitive with other state-of-the-art methods.