AUTHOR=Liao Tingjing , Yang Ruoli , Zhao Peirui , Zhou Wenhua , He Mingfang , Li Liujun 

TITLE=MDAM-DRNet: Dual Channel Residual Network With Multi-Directional Attention Mechanism in Strawberry Leaf Diseases Detection

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 13 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2022.869524

DOI=10.3389/fpls.2022.869524

ISSN=1664-462X

ABSTRACT=Accurate and rapid identification of strawberry leaf diseases is conducive to timely take control measures according to the types of diseases, so as to reduce the losses caused by the diseases. Traditional recognition methods mainly rely on manpower, which requires a lot of time and energy. Developing a more efficient identification method plays an important role in improving the yield and quality of strawberry. Therefore, a detection system for strawberry leaf diseases based on dual channel residual network with multi-directional attention mechanism (MDAM-DRNet) was proposed. (1) In order to fully extract the color features in strawberry disease images, this paper constructs a color feature path at the front end of the network. The color feature information in the image is extracted mainly through color autocorrelation diagram (CAC). (2) In order to fully extract the texture features in the strawberry disease image, this paper constructs a texture feature path at the front end of the network, and mainly extracts the texture feature information in the image through Rotation Invariant Co-occurrence among Adgacent LBPs (RIC-LBP). (3) In order to improve the ability of the model to extract detailed features, in the main frame, this paper proposes an attention mechanism——MDAM. MDAM can allocate weights in the horizontal, vertical and diagonal directions, which can reduce the loss of feature information. In order to solve the problems of gradient disappearance and gradient explosion in the network, the ELU activation function is used in the main frame. Experiments are carried out on our collected database and PlantVillage database. The results show that the network used in this paper has the highest detection accuracy of 98.72%, 96.68%, 95.00%, 95.03%, 96.28% and 96.32% for normal, powder mill, white spot, anthrax and botrytis cinerea. It shows that this method is an effective method for detecting strawberry leaf diseases.