Detecting targets in hyperspectral image under compressive sensing (CS) is crucial for real-time applications in front-ends. However, applying existing deep learning methods directly to compressed hyperspectral image has been challenging, including the uncertainty introduced by the random sensing matrix in the CS acquisition model, and the ineffectuality of traditional similarity metrics. In this paper, we present a solution to these challenges by introducing a Compressive Sensing-based Triplet Transformer Detector (CS-TTD). Our approach involves data augmentation using the Restricted Distribution Property (RDP) to generate samples for a balanced number of positive samples. followed by the Siamese Network with a triplet transformer to transition spectral vectors from the compressed domain to an embedding space. In addition, we introduce a Combine Convolution Net (CCN) classification module to replace traditional metrics for obtaining classification results. To improve the training and make full use of dataset labels, we also present a two-stage training approach, known as intercategory separation and intracategory aggregation (ISIA), combined with Hard-negative mining, Semi Hard-negative mining. Besides, to estimate the minimal number of compressed bands for sufficient detection performance, we introduce a kernel distribution estimation-based sequential iteration (KDE-SI). Our method achieves hyperspectral target detection without reconstruction, and experimental results demonstrate its performance at 0.16 compression ratio is comparative to existing methods in the original domain.

Depth completion aims to recover dense depth maps from sparse depth maps using the corresponding RGB images as guides. Learning guided convolutional network (GuideNet) is one of the state-of-the-art (SoTA) depth completion methods. In this paper, we propose a robust and efficient end-to-end guided spatial propagation network (GSPN), which further improves the effectiveness and efficiency of GuideNet through spatial propagation. Specifically, we expand the receptive field of the content-dependent guided kernels through a spatial propagation network without adding additional parameters. And the resources required by GSPN can be adjusted according to the actual situation. Furthermore, the proposed algorithm can better fuse the information from different sensors, which is one of the main problems of depth completion. We demonstrate the effectiveness of GSPN compared to other SoTA methods on KITTI depth completion and NYUv2 datasets. Our code will be released at https://github.com/scri/GuidedSPN .

Dense depth perception is critical for many applications. However, LiDAR sensors can only provide sparse depth measurements. Therefore, completing the sparse LiDAR data becomes an important task. Due to the rich textural information of RGB images, researchers commonly use synchronized RGB images to guide this depth completion. However, most existing depth completion methods simply fuse LiDAR information with RGB image information through feature concatenation or element-wise addition. In view of this, this paper proposes a method to adaptively fuse the information from these two sensors by generating different convolutional kernels according to the content and positions of the feature vectors. Specifically, we divided the features into different blocks and utilized an attention network to generate a different kernel weight for each block. These kernels were then applied to fuse the multi-modal features. Using the KITTI depth completion dataset, our method outperformed the state-of-the-art FCFR-Net method by 0.01 for the inverse mean absolute error (iMAE) metric. Furthermore, our method achieved a good balance of runtime and accuracy, which would make our method more suitable for some real-time applications.

The relentless growth of artificial intelligence of things (AIoT) applications has tightened sub-milliwatt powerbudget constraints on embedded vision systems. To address this challenge, this brief presents the first brain-inspired intelligent vision sensor (BIVS) based on a novel spiking processing-insensor (PIS) architecture that efficiently converts captured pulsewidth-modulation (PWM) image data into real-time sparse spike features for subsequent intelligent processing, offering an attractive balance between computational accuracy and energy efficiency. A 126 126 BIVS is designed in 180nm CMOS technology and consumes 66.9μJW at 661fps with a resulting efficiency of 5.02 TOPS/W and FoM of 6.37 pJ/pixel/frame. It also achieves a spike feature map accuracy of 99.73% on the MNIST dataset and a classification accuracy of 99.05% when coupled with a simplified spiking convolutional neural network (SCNN).

Dynamic Vision Sensor (DVS) is an event-based imaging technology inspired by biological photoreceptors, which holds great promise for edge computing. The event streams produced by DVS are often contaminated by Background Activity (BA) noise and hot-pixel noise, which degrade downstream processing. Existing filters typically use fixed parameters, resulting in poor adaptability to changing illumination. In this paper, we propose a lightweight Adaptive Event-based Filtering Spiking Neural Network (AEFSNN) to address these limitations. Inspired by homeostatic plasticity, AEFSNN dynamically adjusts neuronal thresholds by monitoring the input-to-output spike ratio, allowing the network to autonomously converge to an optimal operating point across different lighting conditions. Furthermore, we introduce a novel neuronal wake-up mechanism that inhibits processing neurons until triggered by valid input, which effectively suppresses redundant events generated by neighboring activity. Experiments show that AEFSNN is more robust under varying illumination. Compared with current filters, our method increases the Signal-to-Noise Ratio (SNR) of the output data by 1.42–2.33 dB. Additionally, the filtered data improves classification accuracy on downstream tasks, validating its practical value for neuromorphic vision systems.