Research on Classification of Wild Mushroom Based on Feature Fusion and Attention Mechanism

The occurrence of poisoning by edible wild mushroom is frequent, identifying the species of wild mushroom can effectively prevent such incidents. However, due to the wide variety of wild mushrooms and small feature differences between classes, as well as the complex image background in the actual scene, the fine-grained classification accuracy is low. To solve this problem, this paper proposes an improved ResNext50 network, by combining low-level features with high-level features through short connections, the network can learn more useful details from low-level and high-level semantic features. At the same time, in order to reduce the number of parameters of the model, the 3×3 convolution layer of the short connection part adopts depthwise convolution. Transfer learning is used to reduce the training time of the model. Using attention mechanism to reduce the learning of complex background. The experimental results show that compared with the unimproved ResNext50, the number of parameters in our improved ResNext50 is only increased by 2.7M, and the accuracy is 97.94%, which is increased by 1.53%. The comparison results show that compared with DenseNet121, MobilNet-v3, Vgg19 and GoogleNet, the accuracy of the four network models is improved by 7.17%, 3.16%, 2.92% and 2.12%, respectively.