Abstract: Weeding can be performed in the period from 3 to 5 leaves after maize emergence during agricultural production. Among them, chemical and manual weeding still remain the primary operation in maize cultivation. The inter-row mechanized weeding has been ever increasing in recent years. However, some challenges are posed by the intra-row mechanized weeding, due to the limited and discontinuous operation space between plants. In this study, a lightweight intra-row weeding device was designed with the swing-type sine rotary yoke in the maize fields, in order to enhance the weeding efficiency for the less damage rate of seedlings. A coordinate attention (CA) module was introduced before the SPPF layer of the YOLOv5s model. The network was then enhanced to focus on the targets of maize seedlings. The interference of irrelevant information was suppressed, such as the weeds during detection. The accuracy of detection was improved by 0.84 percentage points (94.7%), compared with the original. Better performance was also achieved in the mAP, F1, and P metrics. A single intra-row area was taken as the minimum operational unit. Once the vision system detected an intra-row segment, and then fed back to the control system. The decision was then made to determine whether the area was complete weeding or seedling protection, according to the current speed of the vehicle and intra-row segment. When the intra-row segment fully met the minimum threshold for operation, the control system actuated the motor to rotate. Then the reducer outputted the rotational movement to both sides of the sine rotary yoke mechanism. The push rods were driven to push the weeding ends on both sides for seedling protection and weeding operations. Furthermore, the seedling protection and weeding were directly dominated by the trajectory and absolute speed of the weeding blade. Therefore, a systematic analysis was conducted on the motion trajectory of the weeding blade. Its limit of speed was determined ultimately. A three-factor and three-level experiment was carried out to optimize the operational parameters using the response surface method. The influencing factors included the forward speed of the device, the diameter of the crop protection zone, and the penetration depth of the weeding blade into the soil. The indicators were the weeding efficiency, seedling damage rate, and energy consumption. The results showed that the forward speed of the device shared the most significant impact on the weeding efficiency (P＜0.01), while the diameter of the crop protection zone posed the most significant impact on the seedling damage rate (P＜0.01), and the penetration depth of weeding blades into the soil had the most significant impact on the energy consumption (P＜0.01). The experiment was also conducted to verify the improved model under the optimal combination of parameters, with a forward speed of 0.42 m/s, a crop protection zone diameter of 74 mm, and a weeding blade penetration depth of 18 mm. The better performance of the device was achieved with a weeding efficiency of 92.75%, a seedling damage rate of 2.91%, and power consumption of 129.7 W, fully meeting the requirements for intra-row weeding in the maize field. The observations were made over 15 days after the field trials in the weeded areas. It was found that the majority of the weeds were eradicated, although some weeds were still rooted to continue growing in the soil at the edge of the crop protection zone. The excellent seedling protection and weeding performances were achieved with a low rate of weed recovery. The findings can provide important references for practical applications and machine upgrades.