Recognition-Guided Image Quality Enhancement

Images captured in low light and hazy scenes usually suffer from low contrast,low brightness and serious noise.Adopting these images in the computer vision task often results in poor performance.Therefore,it is necessary to apply high-quality images to the vision tasks,such as target recognition,image classification and semantic segmentation,and image quality enhancement is an effective way to obtain high-quality images.The existing image quality enhancement methods focus on improving the subjective visual quality of images,ignoring the correlation between image quality and recognition task accuracy.Meanwhile,these methods suffer from high complexity and a large number of model parameters.In this thesis,we focus on the design of quality enhancement method to generate high-quality images for the target recognition task.More specifically,we aim to enhance the quality of low-light images and haze images,and the contributions of this thesis are summarized as follows:Appling the low-light images enhanced by using existing methods to target recognition cannot effectively improve the performance,and even results in in lower recognition accuracy.To solve this problem,we proposed an unsupervised low-light image enhancement algorithm based on the feature similarity constraints,which achieves effective enhancement of the features for the recognition task.In the proposed method,the pixel distribution of the image is adjusted by the histogram equalization method.With this adjustment,the features of the image approximate the features of the normal-light image.Based on this adjustment,we firstly constructed a deep image decomposition network to decompose images into the illuminance map and the reflectance map that is more applicable for the recognition task.Then,we eliminated the noise in the reflectance map.Finally,we applied the processed low-light images to the recognition task to guarantee to achieve high recognition performance.The clear images generated by using the existing dehazing algorithms can effectively enhance the performance of the target recognition task,but these methods normally suffer from the high complexity and a large number of model parameters,which cannot satisfy the real-time requirements of the recognition task.To solve this problem,based on the correlation between the multi-scale image features and the recognition accuracy of the target recognition task,we constructed a deep image de-hazing network based on the multi-scale feature fusion and designed a lightweight network by effectively calculating the parameters of the atmospheric scattering model.With the proposed method,we can achieve the real-time quality enhancement for the haze images.In addition,we adopted the proposed multi-scale-related loss function in our method to guarantee that using the enhanced images can achieve good recognition performance.