Accurate emotion classification employing physiological signals provides powerful insights into a person's emotional well-being and is a useful tool to enhance human–computer interaction (HCI) and medical services. Emotion classification relies on electrocardiogram (ECG) and galvanic skin response (GSR) signals and is a developing and fascinating field of study from emotional computing and HCI. ECG measures electrical activity from the heart, but GSR tracks changes in skin conductance connected with emotional arousal. By integrating these physiological signals, it is feasible to infer an individual's emotional state. Machine learning (ML) and deep learning (DL) techniques are executed for analyzing ECG and GSR data to categorize emotions namely stress, fear, happiness, or sadness. These approaches conventionally involve feature extraction in the physiological signals, and then the training of models namely neural network (NN), support vector machine (SVM), or random forests (RF) for classification. Therefore, this study presents a bacterial foraging optimization algorithm with deep learning-assisted emotion recognition and classification (BFOADL-ERC) technique on ECG and GSR signals. The purpose of the BFOADL-ERC technique is to recognize and classify emotions using a hyperparameter-tuned DL model. To achieve this, the BFOADL-ERC technique undergoes data pre-processing where the CWT is applied to transform the frequency domain of ECG and GSR signals into a time–frequency area and thereby derives features. The CWT coefficients are organized to generate a scalogram and are fed into the Inceptionv3 model for the feature extraction process. Moreover, the BFOA-based parameter tuning process can be employed to perfectly choose hyperparameter values of Inceptionv3 technique. Finally, classification of emotions can be performed by the use of an autoencoder (AE). The simulation analysis of the BFOADL-ERC technique is verified on a benchmark dataset and results are investigated below distinct measures. The experimental values inferred improved detection performance of BFOADL-ERC model over other recent approaches.