The standardized and precise identification and detection of seedlings and seeds in rice fields is a prerequisite for achieving the quality detection of mechanical rice planting operations. To address the issues of complex rice field backgrounds, high machinery operation speeds, and difficulty in extracting morphological features during the research on rice planting image recognition, which resulted in low recognition accuracy rates, a lightweight quality detection method based on the improved YOLO v8s was proposed. Firstly, an image acquisition platform for operation quality detection was established through a rice planting quality detection device developed from the Inaka PZ60 type rice transplanter. Images of operation quality were captured to form the ImageSets dataset, and quality detection evaluation indicators were formulated in accordance with relevant national standards. Then by introducing the lightweight GhostNet module, the operational parameters of the network model were reduced. Simultaneously, to enhance the detection performance of the convolutional neural network, the CPCA attention module was incorporated into the detection algorithm, effectively strengthening the feature extraction for the quality of rice planting operations, suppressing the complex background information of the rice field, accurately obtaining the key features of the operation images, and significantly improving the detection effect of numerous small targets such as seedlings and seeds. Secondly, the CIoU loss function in the YOLO v8s model was replaced with the EIoU loss function, enabling the model to have a fast and good convergence speed and localization effect, and achieving precise identification of operation quality. The experimental results indicated that when evaluated using the average precision as the main indicator, the average precision of the improved YOLO v8s model on the test set was 92.41%, with an accuracy of 92.11%, a recall of 92.04%, and an mAP improvement of 7.91, 7.71, 4.28, and 1.03 percentage points, respectively, compared with the YOLO v5s, YOLO v7, YOLO v8s, and Faster R-CNN network models. The detection speed and memory occupancy of the improved model were 88 f/s and 19.2 MB, respectively, which were12.8% and10.7% lower than those of the YOLO v8s model. After tests in the planting environment, it can determine whether the operation quality was qualified, fulfilling the role of quality detection. The improved YOLO v8s network model demonstrated rapid and accurate recognition capabilities for the quality detection of rice field operations, exhibited good robustness, and had remarkable effects in the aspect of rice planting quality detection, providing a detection method for the quality detection of mechanical rice planting.