Kenta Senzaki

Efficient implementation of Filtering-Direction-Controlled Digital TV Filter for image denoising

We propose the efficient implementation of Filtering-Direction-Controlled Digital Total-Variation Filter (FDC-DTVF) for real-time image denoising. FDC-DTVF produces a clear image compared to the conventional pixel-based denoising methods, but has not yet to be considered regarding its efficient implementation. We conduct algorithm optimization by reusing a calculated criterion that is newly introduced in FDC-DTVF, since the criterion is almost invariant in each step of the iterative process. We also conducted a software implementation so we can more efficiently use the CPU resource. The simulation results showed that our implementation achieves the real-time processing for a VGA video, while maintaining the image denoising performance.

CNN-based ship classification method incorporating SAR geometry information

This paper proposes a ship classification method for synthetic aperture radar (SAR) images, which incorporates SAR geometry information into a convolutional neural network (CNN). Most of the conventional methods for ship classification employ appearance-based features. These features extracted from SAR image are not robust to a geometry change because the geometry difference significantly changes the appearance of target objects in SAR images. CNN has a potential to handle the variations in appearance. However, it requires huge training data, which is rarely available in SAR, to implicitly learn geometry-invariant features. In this paper, we propose a CNN-based ship classification method incorporating SAR geometry information. We focus on the incident angle information that is included in a metadata because incident angle change directly affects the appearance of objects. The proposed method enables a network to learn a relationship between the appearance and SAR geometry by utilizing the incident angle information as a condition. Experimental results show that the proposed method improves the classification accuracy by 1.1% as compared to the conventional CNN, which does not utilize incident angle information. Furthermore, our method requires 25% less training data as compared to the conventional CNN to achieve 70% classification accuracy.

Image Clarification Method Based on Structure-Texture Decomposition with Texture Refinement

This paper presents a high quality and low complexity image clarification method, which restores the visibility of images captured in bad weather and poor lighting conditions. A sequential processing of conventional dehazing and backlit correction methods has a problem that textures and noises are overemphasized by the corrections. The proposed method first decomposes a captured image into two components: a structure component forming smooth regions and strong edges and a rest component for fine textures and noises. Image enhancement is conducted based on analyses of the first component, while controlling an amplification factor of the texture component. The utilization of the structure component for the enhancement enables pixel-wise corrections without local area analysis which results in lower computational cost. Experimental results demonstrate that the proposed method can successfully enhance image qualities and its computational cost is reasonable for real-time video processing.