When it is in real-time image acquisition, image blurring caused by lens defects, camera jitter, target movement and so on will result in poor precision of target detection by using the trained deep learning model. Here, a two-stage detection model based on an improved Faster R-CNN and an SSRN-DeblurNet was proposed to perform obstacle detection for blurred farmland images. In the first stage, sharpness evaluation and deblurring were carried out, and the simplified scale recurrent networks (SSRN-〖JP〗DeblurNet) was used for deblurring of blurred farmland images. In the second stage, obstacle detection was implemented by using the improved Faster R-CNN which was added a proposal region optimization network to improve the quality of the regions in the region proposal networks. Then, the proposed two-stage detection model was used to detect eight types of farmland obstacles with self-made blurred dataset. Compared with the original Faster R-CNN, the mean average precision (mAP) value was increased by 12.32 percentage points, and the average detection time of a single image was 0.53s. The results showed that the proposed two-stage model can not only effectively reduce the false detection and missing detection of obstacles in blurred farmland images, but also can meet the real-time detection requirements of tractors operating at low speeds.