When detecting targets in crop images, the detection accuracy of target detection algorithms can be seriously affected due to factors such as dense crop planting and poor imaging quality. In order to optimize the detection performance of crop object detection in YOLOv3, an improved algorithm based on YOLOv3 is proposed. Firstly, the backbone feature extraction network of YOLOv3 is optimized by utilizing the downsampling feature maps outputted by the original network to detect targets, and residual units are added on the basis of the residual blocks in the original network to detect the position information of small crop objects. Moreover, a Gaussian decay function is introduced to attenuate highly overlapping crop candidate boxes in the image, effectively suppressing redundant boxes and reducing false negative rate. Furthermore, the regression loss function is optimized by using CIOU Loss, making the final object localization more accurate during the object detection process. To evaluate the improved YOLOv3 algorithm, a comparative experiment is conducted on a real-world dataset of maize crop images, comparing it with the original YOLOv3 algorithm and the Faster R-CNN algorithm. The results demonstrate that the improved YOLOv3 algorithm can effectively detect small crop targets, exhibiting significantly improved mean average precision and detection speed.