Abstract: Oil spills from thermal power plants can pollute the drainage outlet，if not dealt with in time，it will

also pollute the river environment. Drainage outlet monitoring is from the angle of depression，therefore it is hard to identify accurately with the rectangular detection frame. Based on the image from the depression angle of the

drainage outlet at a power plant，an algorithm of parallelogram detection frame was proposed. The effect of the

characteristics of this parallelogram detection frame and error in angle detection on intersection-over-union（IoU）

was analyzed. Characteristics of two methods for angle detection，i.e. classification and regression were discussed.

Addressing the issue of angle classification，label smoothing based on exponential function was designed. For

detection of the interval angle between categories，the classification- regression detector was designed to study the

offset of angle interval. In the end，the improved classification cross-entropy loss function and IoU-Smooth L1 loss

were introduced for training. According to data concentration tests at the drainage outlet of the power plant，the

proposed method is of the best detection effect.

      Key words: parallelogram dtection frame；power plant drainage outlet；exponential label smoothing；angle

classification-regression detector；environmental protection