In this study, we built a robotic arm control system that combines image processing and deep learning methods for a practical industrial application to grab screwdriver bits on a running conveyor belt. The target objects, that is, the screwdriver bits, have two appearance characteristics: first, the difference between different types mainly lies in the shape of the head, and the head can be regarded as a small object in the image; in addition, the screwdriver bits are in the shape of an elongated cylinder, so that the angles at which they are distributed on the conveyor belt are also different. Small objects can significantly increase the difficulty of object identification, whereas inaccurate grasping due to angle detection errors may cause damage to the products (i.e., screwdriver bits). Under the requirements of accuracy and reliability, we enhanced the existing vision-based robotic arm control system: (1) in the object detection stage, an attention mechanism is added to improve the recognition performance of small objects, and (2) the principal component analysis (PCA) method is adopted to identify the contour of an object in the presence of image noises. Moreover, we combined edge detection to obtain angle information for the gripper’s posture. For these two enhancements, we compared (1) the differences between one-stage object detection methods with and without attention mechanisms and (2) the stability and accuracy of applying PCA to edge detection methods. The experiment data showed that the system proposed in this paper can achieve an object recognition rate of 97.76% and a mean absolute error (MAE) of 0 for angle detection. In experiments consistent with industrial application fields, the system obtains successful placement rates of 97% on a static conveyor belt and 87% on a running conveyor belt.