Deep learning is widely used in electronic nose (E-nose) pattern recognition algorithms. However, deep-learning algorithms are complex in structure and require a large amount of memory resources, inducing difficulty i...
详细信息
Deep learning is widely used in electronic nose (E-nose) pattern recognition algorithms. However, deep-learning algorithms are complex in structure and require a large amount of memory resources, inducing difficulty in the deployment to devices and consequently restricting the practical application of an E-nose. In this work, a lightweight convolutional neural network (LCNN) is designed by weight pruning and quantization and then deployed the E-nose to solve this problem. The whole process including hardware fabrication, data sampling and processing, model design and compression, and deployment to the portable E-nose system has been accomplished. To realize low cost and portability, only a microcontroller unit (MCU) is utilized as the main control chip of the E-nose system to complete model inference and system control. Furthermore, the Kalman filter (KF) is applied to filter out the disturbances during the data acquisition process. In designing a network model for gas concentration prediction, the designed network model consists of only four layers to keep the model size as small as possible, followed by further compression using 80% sparsity weight pruning and dynamic range quantization. The experimental results show that the model after being compressed by nearly ten times still achieves good regression performance with R-2, mean absolute error (MAE), and mean square error (MSE) of 0.9914, 2.932, and 21.052, respectively. Finally, the optimized model was deployed into the MCU to obtain a low-cost and reliable portable E-nose system.
Spiking neural networks are energy efficient and biological interpretability, communicating through sparse, asynchronous spikes, which makes them suitable for neuromorphic hardware. However, due to the nature of binar...
详细信息
Energy trading in distributed microgrids represents an effective means of enhancing the utilization of renewable energy. However, the aggregation of large-scale consumption data may encounter business scalability issu...
详细信息
Convolutional neural networks(CNNs) are widely used in image restoration tasks, but they are limited by local operations and cannot be used to model long-range pixel dependencies. Recently, using the idea of non-local...
详细信息
Convolutional neural networks(CNNs) are widely used in image restoration tasks, but they are limited by local operations and cannot be used to model long-range pixel dependencies. Recently, using the idea of non-local operations, various non-local networks and the Vision Transformer have been proposed to address this problem of CNNs. However, most of these models cannot adaptively process images with different resolutions, and their large number of parameters and computational complexity make them unfavorable for edge deployment. In this paper, we propose an efficient Global Self-Attention Memristive Neural Network (GSA-MNN) for image restoration and present a circuit implementation scheme for GSA-MNN based on memristors. GSA-MNN can both extract global and local information from images and has fully convolutional properties, which can be flexibly applied to different resolution images. Specifically, we design two global attention modules: The Global Spatial Attention Module (GSAM) and the Global Channel Attention Module (GCAM) to complete the modeling and inference of global relations. The GSAM is used to model global spatial relations between the pixels of the feature maps, while the GCAM explores global relations across the channels. In order to deal with image regions with complex textures, we also designed a multi-scale local information extraction module. Powered by these three modules, GSA-MNN enjoys an excellent ability to capture both global and local dependencies in image restoration. Finally, we provide a full-circuit implementation scheme for these three modules, using a modular design to complete the circuit design of the entire GSA-MNN. Benefiting from the programmability of the memristor crossbar, three kinds of image restoration tasks: image deraining, low-light image enhancement, and image dehazing are realized on the same circuit framework by adjusting the configuration parameters. Experimental comparisons with over 20 state-of-the-art methods on 1
Capsule network is a new type of neural network structure. The biggest feature of capsule network is 'vector in vector out', which replaces the previous 'scalar in scalar out'. The pristine capsule net...
详细信息
Encode-decoder structure is used in deep learning for real-time dense segmentation task. On account of the limitation of calculation burden on mobile devices, we present a light-weight asymmetric encoder-decoder netwo...
详细信息
Early automatic and accurate melanoma recognition is an important method to reduce melanoma deaths. Existing methods are less sensitive to the position of the lesion areas. Network training may be affected by the unco...
详细信息
In recent years, there has been a wave of research on English image caption at home and abroad. However, due to the particularity of Chinese image caption task, the research on Chinese image caption has not made good ...
详细信息
Deep learning has shown great advantages in biomedical image segmentation. The classic model U-Net uses a stacked encoding-decoding structure of convolution operations for feature extraction and pixel-level classifica...
详细信息
Images taken in rainy, hazy, and low-light environments severely hinder the performance of outdoor computer vision systems. Most data-driven image restoration methods are task-specific and computationally intensive, w...
详细信息
暂无评论