CDOD-NET: A Crop Disease Object Detection Model Based on Haar Wavelet Network

Current crop disease object detection meth-ods are limited in species scope and performance, making them unsuitable for practical agricultural applications. To address this, building an object detection model based on the YOLOv9c network to detect multiple disease types offers a potential solution. However, the original model's low detection efficiency and the irregular shapes of disease spots further diminish its effectiveness. In order to address these challenges, this paper construct a Crop Disease Object Detection Model based on Haar Wavelet Network (CDOD-NET), and construct the Multi-species Crop Diseases Object Detection (MCDOD) dataset, which includes 64 common crop diseases. CDOD-NET incorporates Arbitrary Kernel Convolution (AKConv) to handle the irregular contours of crop diseases, Haar Wavelet Downsampling (HWD) to reduce spatial resolution while preserving critical information for faster detection, and a Global Attention Mechanism with Convolution (GAMConv) to enhance cross-dimensional fea-ture interactions, improving feature extraction. Experimental results confirm that these methods improve detection performance. On the MCDOD dataset, CDOD-NET surpass state-of-the-art methods, with improvements of 64 f/s in detection speed and 1.0% in accuracy.