Spore detection algorithm of wheat powdery mildew based on weight adaptive feature fusion

Abstract

Aiming at the characteristics of small targets, many interferents and inconspicuous features of spore images of wheat powdery mildew, a weight adaptive feature fusion model is proposed based on SSD network structure to improve the accuracy of spore detection. Firstly, a feature fusion path is constructed to recursively fuse features of various scales from deep to shallow, and at the same time, a layer of feature matrix is added to enhance the utilization of deep and shallow features by the network; Secondly, a hybrid attention module is proposed, which redistributes the weights of features adaptively to enhance the ability of extracting network context information. Finally, the k-means algorithm is used to set the shape of the prior box, which effectively improves the problem that it is difficult to manually adjust the hyperparameter of the neural network. The AP of powdery mildew spores was 91.17%, Compared with the classical SSD detection method, it has been greatly improved.