Photos taken in low-light conditions will suffer from a series of visual problems due to underexposure， such as low brightness， loss of information， noise and color distortion. In order to solve the above problems， a dual-branch residual low-light image enhancement network combined with attention was proposed. Firstly， the improved InceptionV2 was used to extract shallow features. Secondly， Residual Feature extraction Block （RFB） and Dense RFB （DRFB） were used to extract deep features. Thirdly， the shallow and deep features were fused， and the fusion result was input into Brightness Adjustment Module （BAM） to adjust the brightness， and finally the enhanced image was obtained. At the same time， a Feature Fusion Module （FFM） was designed in combination with attention mechanism to capture important feature information， which helped to restore dark areas in low-light images. In addition， a joint loss function was introduced to measure the network training loss from multiple aspects. Experimental results show that， compared with RRM （Robust Retinex Model）， Zero-DCE （Zero-reference Deep Curve Estimation） and EnlightenGAN （Enlighten Generative Adversarial Network）， on LOL （LOw-Light） dataset， the Peak Signal-to-Noise Ratio （PSNR） indicator of the proposed network is increased by 49.9%， 40.0% and 18.5% respectively. Meanwhile， the Structural Similarity Index Measure （SSIM） indicator of the proposed network is increased by 20.3%， 50.0% and 34.5% compared with those of the above three on LOL?V2 dataset. The proposed network improves the brightness of low-light images while reducing noise， color distortion and artifacts， resulting in sharper and more natural enhanced images.

Aiming at the problem that the traditional pooling algorithms in Convolutional Neural Network （CNN） cannot well consider the correlation between each element in the pooling domain and the features contained in the pooling domain， a pooling algorithm based on Gaussian function was proposed. Firstly， according to the value of each element in the pooling domain and the maximum value of all elements， the three parameter values of the Gaussian function were calculated. Then， the Gaussian function was used to calculate the weights of all elements in the pooling domain. Finally， the weighted average value of all elements in the pooling domain was calculated according to these weights. Finally， the obtained value was used as the pooling result. LeNet5， VGG （Visual Geometry Group）16， ResNet （Residual Network）18 and MobileNet v3 were selected as the experimental models. Experiments were carried out on public datasets CIFAR-10， Fer2013 and German Traffic Sign Recognition Benchmark （GTSRB）， and max pooling， average pooling， random pooling， mixed pooling， fuzzy pooling， fused random pooling and soft pooling were selected to compare. Experimental results show that the proposed algorithm improves the accuracy by 0.5 percentage points to 6 percentage points compared with other algorithms on the three datasets， and the running efficiency of the proposed algorithm is higher than those of the other pooling algorithms except max pooling algorithm and average pooling algorithm， so as to verify that the proposed algorithm is effective and suitable for the situations where the operation time demand is not high but the accuracy demand is high.