In order to improve the detection performance of YOLOv4 object detection algorithm for small apple leaf lesions, a PSA-YOLO network with low computational cost and high accuracy was proposed, which integrated a Focus layer and the pyramid squeeze attention block in the CSPDarknet, and the strategy of network depth reduction was adopted. Finally, the PSA-CSPDarknet-1 was built on the basis of CSPDarknet53. The experimental results showed that the computational complexity of PSA-CSPDarknet-1 was reduced by 30.4% compared with the CSPDarknet53 and the detection accuracy of the network for small lesions (covering area less than 32 pixels×32 pixels) was improved by 2.9 percentage points. In the neck, a spatial pyramid convolution and pooling module was built to enhance multi-scale information extraction in spatial dimensions with a small computational cost, and α-CIoU loss function for the bounding box was used to improve the detection accuracy of bounding boxes for improving the detection accuracy of lesions under the high IoU threshold. According to the experimental results, the proposed PSA-YOLO network achieved 88.2% AP50 and it achieved 49.8% COCO AP@[0.5∶0.05∶0.95] in the apple leaf lesion dataset, which was 3.5 percentage points higher than that of YOLOv4. At the same time, the feature extraction ability of the network for small lesions was more improved, and APS was 3.9 percentage points higher than that of YOLOv4, respectively. The detection speed on a single NVIDIA GTX TITAN V reached 69 frames per second, which was 13 frames per second faster than that of YOLOv4.