CBAM-DCE: A Non-Reference Image Correction Algorithm for Uneven Illumination

Abstract

Affected by the change in daytime illumination sequence and by the shooting angle in the complex field environment, the kiwifruit images possess the unfriendly features of uneven illumination, such as local darkness and local brightness. The ill-posed image with uneven illumination will seriously constraint the subsequent image analysis processing. Current deep learning methods have achieved satisfactory results, and a large number of paired images (one is the input image, one is the ground truth image) is required to train the better network performance. However, it is difficult to capture ground truth images of the kiwifruit in the field. Based on this, the paper proposed Convolutional Block Attention Module Deep Curve Estimation (CBAM-DCE) to accomplish a non-reference illumination unevenness correction for field kiwifruit images. A deep learning network model is used to estimate the image-specific curve for image enhancement, and a non-reference loss function is applied to evaluate the image enhancement effect. Compared with seven related enhancement algorithms, the presented algorithm shakes off uneven illumination or normal-light image pairs for training. Five different public datasets and our Kiwifruit dataset were used in the experiments. Experiments demonstrate that our proposed CBAM-DCE is superior to other state-of-the-art algorithms for enhancing natural images under different lighting conditions.