The widespread application of deep learning in the agricultural field has advanced radish segmentation and detection technologies. However, the issue of high dynamic range in backlit images captured by smartphones affects the monitoring of radish growth. Deep learning-based methods tend to perform poorly in complex lighting conditions, struggling with background interference, overexposure, and information loss. In this paper, we propose a new DoG-Mask RCNN network based on JND (Just Noticeable Difference) visual contrast compensation preprocessing. The network incorporates data preprocessing techniques for challenging environments as input, where the DoG layer is used to suppress overexposure in special conditions and filter out irrelevant background information. This layer retains spot details through its filtering capabilities, while the increment layer ensures the integrity of the feature information. We conducted experiments on radish mixed with traditional crops in different challenging environments to demonstrate the effectiveness of our method in detecting radish parts under various conditions. Compared to the latest monitoring methods, our approach shows better detection stability.