As one subject for the use of the object detection to Bakanae-disease detection, this article makes two contributions and get impressive results based on famous Faster-RCNN network. To make Faster-RCNN effective and robust, the author introduces the deformable convolution attention mechanism and feature pyramid network, the author chose Faster-RCNN as baseline, and then proposed the following two improvements based on the characteristics of the rice seedlings. Considering the arbitrary shape of

more »

... rice seedlings, the Deformable Convolution is introduced in the backbone network to capture the characteristics of targets in any direction. As for the redundant information in the network, a Channel Attention module is added after the FPN to adaptively strengthen the useful information and suppress the interference of redundant information. The two improvements have greatly improved the learning ability of the network, which is particularly important for the detection of irregularly shaped and undefined targets in rice seedlings. In a word, the author designed a strong and reliable network for Bakanae-disease detection and actually tested it.