Automatic gear defect detection equipment is relatively expensive, so small and medium-sized enterprises cannot afford the cost of such equipment. Therefore, most companies still use manual inspection methods for gear defect detection. Manual inspection methods not only take a long time but also has uneven detection quality. This paper proposes to use AI technology to build a cheap and fast gear defect detection method. And this method is used to complete the detection of gear tooth profile defects, tooth pitch defects and central hole defects. The method proposed in this paper is divided into four steps. In the first step, the ResNet model [1] is used to classify whether the gear image is complete or not. In the second step, the YOLOv4 model [2] is used to find the rectangular area of the tooth shape and tooth pitch in the image and cut it out. The third step is to use the UNet model [3] to segment the tooth profile and pitch profile, and calculate the area occupied by the profile. Finally, whether the difference from the average area is too large is used as the basis for judging whether the gear is defective. In the experiment result, 186 gear images are used for detection, and the obtained accuracy is about 91%. This result in addition to verifying the feasibility of the proposed method, it is also found that the proposed method can quickly and accurately detect gear defects that are difficult to judge by human eyes.

Many people believe that when drowning occurs, there will be calls for help. In fact, people who are drowning do not get too many splashes or cry for help. They only try to get themselves out of the water by treading on the water. The drowning condition may cause serious brain damage, so it is extremely important to shorten the time it takes to detect the occurrence of drowning and rescue.This paper proposes using computer image processing technology to introduce artificial intelligence motion technology, mounting the camera on the bottom of the swimming pool, and use OpenPose to mark the image joint point features, and input the captured joint point features into the recursive neural network to determine whether the swimmer is drowning. The final training result is about 89.4% accurate, so it can be used to assist on-site lifeguards to detect swimmers who may be drowning, and to reduce incidents that cannot be detected immediately