A precise apple leaf diseases detection using BCTNet under unconstrained environments

Accurate detection of apple diseases is crucial for agricultural production as it enables timely diagnosis and control, thereby reducing crop yield losses. However, detecting apple leaf spots presents significant challenges due to unconstrained environmental conditions and multi-scale variations. This paper proposes a novel deep learning-based detection algorithm to accurately detect multi-scale apple leaf spots in unconstrained environments. The proposed method consists of several steps. First, a dataset of apple leaf disease spots with high-quality labels is created under the guidance of agricultural experts. Second, a Bole convolution module (BCM) for reducing the interference of redundant feature information on extracting feature information of apple leaf disease images in unconstrained environments is designed. Third, a cross-attention module (CAM) for reducing the computational effort of the detection network in non-diseased regions to reduce the impact of background interference information on the feature representation capability of the network is proposed. Fourth, to reduce the loss of surface and deep feature information of apple leaf diseases during feature fusion and communication, we employ a bidirectional transposition feature pyramid network (BTFPN) to address this problem. Finally, the proposed apple leaf disease detection network, which combines the Bole convolution module, Cross-attention module, and Bidirectional Transposed feature pyramid Network (BCTNet), achieves an accuracy of 85.23% and an average detection speed of 33 FPS on the self-built dataset. The proposed method outperforms other state-of-the-art methods in terms of accuracy. From visible light images, it can detect multi-scale disease spots on apple leaf surfaces in natural environments and provide decision information to growers and pesticide spraying robots. The proposed method can replace traditional manual diagnosis methods, optimize the spraying efficiency of pesticide robots, and reduce pesticide waste in non-diseased areas. Part of the dataset used in this paper can be found in the https://github.com/ZhouGuoXiong/BCTNet.