Wild Mushroom Image Segmentation Algorithm Based on Improved DeepLabV3+

Wild mushrooms in China are diverse and widely distributed, particularly in southwestern regions like Yunnan and Guizhou, where many inedible species pose a risk of poisoning upon accidental ingestion. To mitigate these incidents, researchers are increasingly focused on species recognition. However, the natural growth environments of wild mushrooms introduce significant background disturbance, complicating accurate identification. Consequently, separating the background from the image is crucial for enhancing recognition accuracy. To solve the above problems, this paper takes the common wild mushrooms in Guizhou province as the research object and proposes a deep learning network based on the improved DeepLabV3+ for the background segmentation task of wild mushroom images. Firstly, the network Xception backbone feature extraction network of the original DeepLabV3+ is replaced with MobileNetV2 network, which makes the network more lightweight and suitable for deployment into mobile devices. Secondly, the Feature Pyramid Networks feature fusion branch is proposed to realize the fusion of different levels of feature maps, so that the network can extract richer wild mushroom features and improve the wild mushroom image segmentation accuracy. And then, a multi-layer cascade superposition is used to replace the original Atrous Spatial Pyramid Pooling module with Dense Atrous Spatial Pyramid Pooling as an enhanced feature extraction network to increase the perceptual field of the network and obtain the multi-scale information of the wild mushroom image. Finally, the ResNet-18 network was used to identify the segmented images. Experimental results show that the improved DeepLabV3+ has better segmentation performance. The average segmentation accuracy of the algorithm is 96.81%, the mean Intersection over Union is 92.36%, and the average accuracy after ResNet-18 recognition reaches 93.8%. In the comparison experiments, our segmentation algorithm has a higher segmentation accuracy and a lower model size. This method can provide technical support for automatic segmentation and identification of wild mushroom images in natural environment.