The infrared small-target lacks effective information such as shape and texture, so it is difficult to detect small-target effectively. In order to solve this problem, a new deep learning network is proposed: Infrared Small-target Detection U-Net (ISTDU-Net). ISTDU-Net is a deep learning network based on U-shaped structure. It converts a single frame infrared image into a target probability likelihood map of image pixels. ISTDU-Net not only introduces feature map groups in network down-sampling, sensing and enhancing the weights of small target feature map groups to improve the characterization ability of small targets; but also introduces a fully connected layer in jump connection to suppress a large number of backgrounds with similar structures from the global receptive field, thus improving the contrast between targets and backgrounds. Experimental results show that the ISTDU-Net proposed in this letter can detect small infrared targets in complex backgrounds. Compared with other algorithms, ISTDU-Net has a better ROC curve with low false alarm rate, which the AUC value is 0.9977.

In traditional imaging system design, the optical lens is often optimized toward the artificial optimization target like modulation transfer function and field-of-view (FoV). This usually leads to complex stacks of lenses. In order to reduce the complexity, we propose an end-to-end single lens imaging system design method. First, the imaging and processing model is established, whose input end is the ground truth image, and the output end is the restored image by Res-Unet. Then, with the optimization target of minimizing the difference between the restored image and the ground truth image, the parameters of the lens surface and the parameters of the restoration algorithm are optimized simultaneously by deep learning. In order to realize the end-to-end design, the imaging model is required to be differentiable to the lens parameters, so a fast differentiable ray tracing model is proposed. A single lens imaging system with high-quality large FoV (47°) has been designed by the end-to-end method. This method will have a wide application prospects in the design of light and small optoelectronic imaging systems.

End-to-end single-lens imaging system design is a method to optimize both optical system and reconstruction algorithm. Most end-to-end single lens systems use convolutional neural networks (CNN) for image restoration, which fit the transformation relationship between the aberrated image and the ground truth image in the training set. Based on the principle of optical imaging, we realize non-blind image restoration through Wiener deconvolution. Wiener deconvolution is improved with the powerful fitting ability of depth learning so that the noise parameters and the blur kernel in Wiener deconvolution can be simultaneously optimized with the optical parameters in the lens. Extensive comparative tests have been conducted to demonstrate the single-lens imaging system obtained by our method has more stable imaging quality and a 40 times greater imaging speed than the method using CNN restoration algorithm.

Infrared small target detection is crucial for infrared search and tracking systems. Saliency-based detection techniques have gradually demonstrated their superiority. Existing methods often adopt a fixed rectangular window for sliding window filtering to measure the saliency of suspected targets, making it difficult to accommodate targets of different sizes and easily leading to contamination by non-target pixels. Therefore, the rectangular sampling window approach increases the difficulty of detecting infrared small targets. In this letter, a novel flexible window local contrast measure (FWLCM) is proposed. Firstly, a flexible window estimation method is suggested that dynamically adjusts the shape of the saliency measurement window based on the heterogeneous variations between the target and the background by performing a flexible region growing. Especially, the local contrast is calculated at an adaptive scale to enhance the target saliency. Secondly, a directional sector local contrast calculation method is introduced to optimize the statistical way of background region characteristics, reducing the interference of discrete noise in the background statistical features. Finally, adaptive threshold segmentation is utilized to extract the real target. Compared with existing methods, experimental results on four real datasets demonstrate the superiority of FWLCM in complex backgrounds.

The restoration algorithm can reduce the aberration blur in the single-lens computational imaging system. The single lens and the restoration algorithm should be co-designed to improve the imaging performance of the whole computational imaging system. By investigating the influence of aberrations on computational imaging quality, we find there are aberrations that cannot be removed by restoration algorithm, which cause the loss of high-resolution information. However, such aberrations are not constrained in existing co-design methods. To solve this problem, we propose the generalized aberration regularization (GAR), which mainly constrains the aberrations that reduce the modulation transfer function (MTF) at the high frequency. Experimental results demonstrate that our approach can significantly improve the quality of high-resolution information in the recovery image compared with the existing co-design methods.