Objective To improve the efficiency and accuracy of walnut kernel sorting in existing food production lines.Methods Based on the intelligent production line for walnut kernel sorting, an improved EfficientDet model is proposed for the intelligent sorting of walnut kernels in food production lines. A convolutional attention mechanism module is introduced into the backbone network to strength the ability of the model to focus on food regions. The bidirectional feature pyramid network is improved to enhance the detection ability of the model for different scales of food. The original activation function is optimized through Dynamic ReLU activation function to enhance the detection performance of the model for food, and the optimized model is deployed in food production for experimental verification.Results The experimental method achieves precise recognition and efficient classification of normal, broken shells, black spots, and dried walnut kernels in the walnut kernel sorting task. This method achieves the detection of a single image within 18 ms, with the average accuracy of 97.92% and a false detection rate reduced to 1.0%. This can effectively improve the automation level of the food production line.Conclusion This intelligent sorting method effectively solves the problems of low efficiency and poor accuracy of conventional sorting methods, and has good application prospects and promotion value in food production line automation.