Imaging is very important in many industries, however, it often suffers from low light and colour shift issues. For tasks that require object detection, quality issue can reduce model accuracy. This work aimed to enhance the performance of few object detectors such as Mask R-CNN, YOLOv5, and YOLOv8 on low-quality images using a combined histogram matching and model stacking techniques. We used 2 datasets which exhibit image quality issues, namely Roboflow100 underwater object and oil palm fruit (Elaeis guineensis) images collected from a plantation in Indonesia. The images were processed through two distinct gates where two models were trained on original images ( \(M_1\) ) and on histogram matched images ( \(M_2\) ). The output from the two gates were stacked and applied Non-Maximum Suppression (NMS) to reduce the overlapping. The predicted class from the second model shows good agreement with ground truth due to colour correction by histogram matching (0.651 F1-score). Hence, we modified the NMS to assign the class from \(M_2\) and presented final result. The result demonstrated a notable improvement in accuracy up to F1-score of 0.811, confirming the effectiveness of the proposed method in enhancing object detection performance on low-quality images.