PAL-net: a light correction method for pears based on image information characteristics

• Signal processing analysis was performed on images with different lighting effects; • A correction model based on a CNN architecture was designed.; • Significantly enhance the model's object detection capabilities at night; • Provide visual technology support for drones and robots to operate at night. Complex lighting conditions often lead to reduced accuracy in pear image object detection. Addressing this research gap, this paper first quantitatively analyzes the varying impacts of lighting at different times of day on pear object detection performance. Based on the identified variations and Retinex theory, a convolutional neural network deep learning model is constructed. This model enhances and restores pear image feature information through light correction, thereby improving detection accuracy—marking the first application of deep learning light correction technology in pear image detection. Experimental results demonstrate that after processing by this model, the mean average precision (mAP) for pear object detection increases by 9.6–15.1 percentage points. To further optimize model practicality, knowledge distillation was employed to achieve model lightweighting, yielding a lite version. Comparative experiments demonstrate that the lite model outperforms the base model in both evaluation metrics and processed image detection results: it further improves the mAP for pear detection, reaching a maximum of over 25 percentage points, while reducing the parameters to less than one tenth of the basic model, only 2.4M, and meeting the real-time requirements of agricultural operations. This study provides a high-precision, lightweight technical solution for detecting agricultural products under complex lighting conditions, demonstrating clear practical value.