With the rapid development of deep learning, there have been numerous object-detection algorithms based on deep learning. However, during the study of the YOLOv5 model, it was found that the SPP-based feature extraction structure may lose some details during the max-pooling process. Additionally, the LeakyReLU activation function lacks sufficient smoothness, and the multi-scale feature fusion using $\mathbf{FPN+PAN}$ structure in its Neck part is not sufficient. To address these issues, this paper proposes an improved YOLOv5 model based on the pyramid structure of dilated convolutions, which introduces dilated convolutions with channel attention mechanism based on the SPP structure, and makes some adjustments in activation function and Neck part. Comparative experiments on the PASCAL VOC2007 and VOC2012 benchmark datasets show that the proposed model achieves higher detection accuracy.
Images captured under low-light conditions often suffer from inadequate lightness and low color contrast, resulting in reduced performance of computer vision-related applications. In this work, we propose an unsupervised network for enhancing low-light images, which enhances the lightness and color constancy of images. To achieve this, we introduce two sub-modules, LC-Net and GP-Net, which estimate high-order curves for lightness enhancement and gray pixels for color constancy, respectively. The proposed method shows promising results in improving the quality of low-light images. To further enhance the information reasoning capability of our proposed method, we adopt an attention block that combines channel and space attention as a basic unit for the network. We evaluate the performance of our method quantitatively and visually in comparison to other existing low-light image enhancement techniques through experiments, the results indicate our method can outperform other approaches greatly in terms of color fidelity and lightness enhancement.
The diagnosis of colorectal cancer by endoscopy is subjective and highly dependent on the physician's experience and knowledge. Therefore, it is vital to develop a diagnostic support system to reduce the variation in diagnostic results among physicians and improve diagnostic accuracy. Our interviews with physicians revealed that physicians pay attention to blood vessels. This paper proposes a feature extraction method based on blood vessels using a vascular enhancement method and a lifting dyadic wavelet transform. This method facilitates the acquisition of blood vessel features by reconstructing the high-pass filter of the dyadic wavelet transform so that the high-frequency component in the blood vessel area after the blood vessels are enhanced by the blood vessel enhancement method. In classification experiments using the features obtained by the proposed method, we confirmed that the classification performance was superior to existing feature extraction methods. The classification experiments using the features obtained by the proposed method showed that it outper-forms existing feature extraction methods. These results indi-cate that the blood vessel-based features are beneficial in detecting early-stage colorectal cancer.
In this work, we introduce the two-sided Quaternion Fourier transform and main properties. Using the decomposition of quaternion, we propose a new form of Plancherel's formula related to transformation. We also obtain a form of convolution theorem associated with the two-sided Quaternion Fourier Trans-form.
In many methods, before inputting Electroencephalography (EEG) signals into the Convolutional Neural Network (CNN), the signals are usually preprocessed by low-pass filter, wavelet threshold denoising and so on. In this paper, it is found that these operations can be completely replaced by CNN, which automatically ignores noise interference in the recognition training process. A light and effective network named Shallow Residual CNN (SRCNN) is proposed to recognize the EEG hand-motion signals whose dataset is from Kaggle competition. The experiment results indicated that our proposed method can directly manage the raw data and have a better performance under ROC curve area evaluation when comparing with the classic deep learning network ResNet 34. This paper can be a good example to extend the CNN-based scheme to more types of EEG signal.
