Stink bug are one of the major causes of Longan damage - a crop of considerable economic value that is extensively cultivated in Vietnam. Early detection of this insect pest enables growers to implement effective control measures, thereby enhancing fruit quality and yield. Current object detection approaches have yielded incomplete results in insect pest detection because they are typically small, obscured, and vary in shape in the images. To tackle this challenge, we propose a novel approach based on YOLOv7 to improve the ability of stink bug detection. In the proposed model, a feature attention module is invented and plugged into the head section of the baseline detector to yield useful features for prediction task. In addition, an extra prediction layer is introduced to enhance the performance of small instances detection. Experimental results demonstrate that our method achieved the highest object detection performance, with an average precision (AP) score of 78.21% on a collected dataset, surpassing state-of-the-art object detection models.