Computer vision is one of the commonly used technical methods in the field of cotton seed detection. It has been widely used in the field of non-destructive inspection of agricultural products. However, in most cases, it is difficult for researchers to use computer vision to detect small-sized objects such as cotton seeds on both sides. The detection effect is not good. Aiming at this problem, a type of cotton seed detection and sorting device was designed, which used the transparent characteristics of the acrylic plate under strong light and white background to slide the cotton seed into the groove of the transparent acrylic plate through the feeding device. With the rotation of the turntable, the front and back images of the same batch of cotton were collected by two CCD cameras at different positions. The improved YOLO v4 target detection algorithm was used to detect damaged cotton seeds. The experimental results showed that the model established by this method can detect damaged and intact cotton seeds in the population cotton seeds with an accuracy of 95.33%, recall rate of 96.31%, and missed detection rate of 0. The detection effect was better than that of the original YOLO v4 network, respectively. The proposed method realized the identification of the damage of double-sided group cotton seed, and provided technical support for the subsequent research and development of related delinted cotton seed intelligent detection equipment.