Abstract: Abstract: A traceability system is able to provide an opportunity of obtaining quality and safety information of agricultural products from farm to table. Over the past years, many researchers have engaged in it and continuously made significant progresses, however, the identification of fruit has not been resolved perfectly. Traditional identification technologies include barcode, QR (quick response) code and RFID (radio frequency identification) card. Unfortunately, the barcode and QR code paper tags are easily stained in fruit storage and transportation with high humidity, while RFID is rather expensive for low value products like fruits. Biometric identification is distinctive, and measurable characteristics are used to label and describe individuals. The physical characteristics and traits used in biometric identification include but are not limited to fingerprint, iris, voice and face. Since biometric identifiers are unique to individuals, they are more reliable in identification than traditional methods. As a newer and safer technology, biometrics is being extensively studied and widely used in human identification currently. In recent years, it is also used on livestock such as cattle and sheep, yet it has not been reported so far on biometric identification of fresh fruit. This study tried to introduce biometric technology into fruit identification of traceability system to fill in this gap. Based on the algorithm of iris recognition, a watermelon identification method was proposed to exploit texture information of the area around the fruit pedicel. The method could be briefed as the following procedures. The color image of watermelon was firstly transformed into gray image to reduce computational complexity and improve calculation efficiency. Then 2 concentric circles were constructed to center the watermelon pedicel, which would be the annulus area for extracting texture information. To ensure the invariance of translation and scaling, every original annulus area image was normalized to the same size using polar coordinates transformation. The texture information of normalized rectangle image was filtered using one-dimensional Gabor filters to extract the orientation and phase information due to its excellent direction selectivity and frequency selectivity. The identification of watermelon texture code based on the above steps is a typical pattern matching problem. This study identifies watermelon using the Hamming distance between texture codes. When the Hamming distance is less than the preset threshold, it would be judged to be from the same individual, vice versa. To eliminate the effect of image rotation, this study converts the rotation of the texture image into code offset, and moves forward or backward the code to overcome the problem of image rotation. The experiments were carried out to verify the feasibility and effectiveness of identification method using the texture of annulus area around the watermelon pedicel. As a watermelon can usually be stored for 2 weeks without spoiling at room temperature, we acquired an image for each of 100 watermelons on the harvest day, the 7th day and the 14th day after harvest and figured out 3 groups of texture codes for subsequent phases. The Hamming distances were computed between each texture on the 7th day and on the harvest day, as well as the 14th day and the harvest day. We can get the following results from the experiments. Firstly, every watermelon has its unique texture code. In other words, the texture code of any watermelon is similar with the one from the same individual, while significantly different from the one from any other watermelon. Secondly, we compared 2 groups of Hamming distances from different time using paired-samples T-test and figured out P-value that was 0.8264. It showed that there was no significant difference between the 2 groups of Hamming distances above, which implied the texture feature of watermelon was relatively stable within 2 weeks after harvest. Thirdly, we divided the 100 watermelon into 2 equal groups and calculated the threshold using one group, and then discriminated the other group with the threshold. All the Hamming distances were classified with 100% accuracy rate and 100% recall rate. Finally, the pictures of watermelon would not be exactly identical if they were acquired at different time or on different conditions. These variations likely cause translation, scaling and rotation of watermelon in an image, and the experiment shows that the method is adaptive to these inevitable variations. Although some problems need to be studied further, this research provides a new idea of biometric identification for fruit quality and safety traceability system.