Low-Light Image Reconstruction Based On Improved Convolutional Autoencoder

Machine vision target detection can be mainly divided into traditional detection methods represented by Hof circle detection and template matching, and detection methods based on deep learning. Traditional target detection methods need to de-noise the image, binarization and other pre-processing methods to ensure that the contour of the target can be extracted, but in the actual operating environment, the interference generated by factors such as light and the surrounding environment often leads to the contour of the target in the image is difficult to be extracted, reducing the success rate of the traditional methods to extract the circle in the complex environment. And the deep learning based circle detection method can solve this problem. Deep learning methods can obtain more specific features, have a higher degree of recognition of the learned features, and have improved performance over traditional machine learning methods. Therefore, we propose a convolutional autoencoder-based low-light image processing method, which improves the loss function by introducing the luminance module of SSIM in order to realize the adaptive enhancement of the luminance of low-light images.