DynaFogBerryNet: A unified model for image dehazing and object detection toward accurate strawberry maturity monitoring

Convolutional neural network (CNN)-based strawberry maturity detection algorithms typically perform well on standard datasets. However, under adverse weather conditions such as fog and haze, image quality deteriorates markedly, resulting in a substantial drop in detection performance. To address the issue, this paper proposes DynaFogBerryNet, a strawberry maturity detection model tailored for foggy environments, to improve detection by integrating image dehazing with object detection. The model combines a dehazing module and a YOLOv8 object detection module, leveraging the advantages of both CNN and Transformer networks and introducing a brightness correction mechanism to remove impurities and blur from foggy images. In YOLOv8, dynamic convolution residual module and ultra-lightweight dynamic upsampling module are incorporated to alleviate the problem of fine-grained detail loss in foggy images. Through an attention-driven fusion strategy, the model integrates the enhanced features from the dehazed network into the detection pipeline, significantly improving object detection performance and enabling reliable strawberry ripeness assessment under foggy conditions. To assess the performance of the proposed model, a dedicated foggy strawberry image dataset is created, and the network is trained and evaluated using both synthetic and real-world datasets. Experimental results demonstrate that DynaFogBerryNet effectively suppresses halo artifacts in low-light foggy images via dynamic convolution kernel adaptation, achieving a 0.5% improvement in mAP for strawberry maturity detection compared to YOLOv8, with particular enhancement in small target recall, showing an increase of 1.19 percentage points. • Propose an effective unified framework for joint image dehazing and object detection. • Design DynamicBnAct and DySample modules to enhance fine-grained feature extraction. • Construct a foggy strawberry dataset to improve model adaptability and generalization.