This paper presents an evolutionary synthesis of feature extraction programs for object recognition. The evolutionary synthesis method employed is based on linear genetic programming which is combined with redundancy-removed recombination. The evolutionary synthesis can automatically construct feature extraction programs for a given object recognition problem, without any domain-specific knowledge. Experiments were done on a lawn weed detection problem with both a low-level performance measure, i.e., segmentation accuracy, and an application-level performance measure, i.e., simulated weed control performance. Compared with four human-designed lawn weed detection methods, the results show that the performance of synthesized feature extraction programs is significantly better than three human-designed methods when evaluated with the low-level measure, and is better than two human-designed methods according to the application-level measure.