To achieve early detection of crop diseases, a crop disease early detection model was proposed based on infrared thermal imaging and improved YOLO v5. The CSPD-arknet was used as the main feature extraction network, and the YOLO v5 stride-2 convolution was replaced by the SPD-Conv module, which were respectively the five stride-2 convolution layers in the main network and the two stride-2 convolution layers in the Neck. This can improve its accuracy while maintaining the same level of parameter size and outputting three different scales of feature layers in the downstream stage. In order to enhance the interdependence between modeling channels, channel feature responses were adaptively recalibrated and SE mechanism was introduced to enhance feature extraction ability. In order to reduce model calculation and improve model speed, SPPF was introduced. After testing, the improved YOLO v5 algorithm had the best detection performance with an mAP of 95.7%, which was respectively 4.7 percentage points, 8.8 percentage points, 19.0 percentage points, and 3.5 percentage points higher than that of YOLO v3, YOLO v4, SSD, and YOLO v5 networks. Compared with the improved network before improvement, it also improved the detection of crop diseases under different temperature gradients. The mAP of five gradients were 91.0%, 91.6%, 90.4%, 92.6%, and 94.0%, which were higher than those before improvement by 3.6 percentage points, 1.5 percentage points, 7.2 percentage points, 0.6 percentage points, and 0.9 percentage points, respectively. The size of the improved YOLO v5 model was 13.755MB, which was lower than 3.687MB of the basic network before the improvement. The results showed that improving YOLO v5 can accurately and quickly detect early diseases, which can provide certain technical support for the development of early disease detection instruments.